Automating Assessment and Providing Personalized Feedback in E-Learning: The Power of Template Matching

Abstract

This research addressed the need to enhance template-matching performance in e-learning and automated assessments within Egypt’s evolving educational landscape, marked by the importance of e-learning during the COVID-19 pandemic. Despite the widespread adoption of e-learning, robust template-matching feedback mechanisms should still be developed for personalization, engagement, and learning outcomes. This study augmented the conventional best-buddies similarity (BBS) approach with four feature descriptors, Harris, scale-invariant feature transform (SIFT), speeded-up robust features (SURF), and maximally stable extremal regions (MSER), to enhance template-matching performance in e-learning. We systematically selected algorithms, integrated them into enhanced BBS schemes, and assessed their effectiveness against a baseline BBS approach using challenging data samples. A systematic algorithm selection process involving multiple reviewers was employed. Chosen algorithms were integrated into enhanced BBS schemes and rigorously evaluated. The results showed that the proposed schemes exhibited enhanced template-matching performance, suggesting potential improvements in personalization, engagement, and learning outcomes. Further, the study highlights the importance of robust template-matching feedback in e-learning, offering insights into improving educational quality. The findings enrich e-learning experiences, suggesting avenues for refining e-learning platforms and positively impacting the Egyptian education sector.

1. Introduction

Online learning has become increasingly popular as a complement to conventional face-to-face education [1,2,3]. This growth is partly because students are active technology users and often use it more than needed. According to Felix [4], examining educational institutions and their strategies in the 21st century reveals several prominent characteristics. These include adaptability, inclusivity, collaboration, authenticity, relevance, and expanded institutional boundaries. Considering the current educational strategies naturally leads to a social constructivist paradigm for learning and teaching, emphasizing the importance of active participation and social interaction. Despite some challenges, a combination of social and cognitive constructivist activities can optimize time usage, incorporating personalized intelligent computer-assisted language learning (ICALL) systems. This hybrid approach improves learning outcomes while efficiently managing the time constraints of addressing the changes regarding effective education strategies.

Consequently, online learning is preferred for its flexibility compared to conventional educational settings. To support e-learning, various platforms have emerged over the past few decades. These online platforms have experienced a surge in use due to their significance as educational support tools with their resources, such as chat, forums, and wikis, facilitating communication between instructors, students, and course content [5]. Nevertheless, one of the main challenges instructors face in online learning environments is maintaining track of an indefinite number of students [6]. Tracking the students further affirms their availability, marks their attendance, and ensures feedback from them as required by the instructor [7]. Notably, feedback is an essential part of the teaching–learning process, enabling students to identify gaps in their knowledge and estimate their learning progress [8]. Productive feedback provides specific information that helps fill the gap between expected and actual understanding of content or abilities. Feedback allows students to work on areas with poor knowledge or skills hindering their learning progress [9]. Research has shown that students’ feedback positively affects learning outcomes. For example, it is witnessed that feedback significantly enhances student learning and satisfaction [10]. Existing research [11] has proposed 12 main conditions that support effective feedback, emphasizing the significance of carefully designing feedback strategies organized into three categories: capacity, projects, and culture.

Similarly, personalized feedback in e-learning is tailored to each learner’s specific needs and performance [12]. In an e-learning context, personalized feedback is generated automatically by a computer-based system or provided by an instructor or mentor [13]. Personalized feedback in e-learning supports learners in improving their understanding of the material and their performance in assessments. This feedback can take many forms, such as detailed explanations of correct and incorrect answers, recommendations for further study or practice, and incentives to persist in learning. Personalized feedback can also consider each learner’s unique learning styles and preferences, providing guidance and resources that are more effective for their individual needs [14]. This can help to increase motivation and engagement, leading to better learning outcomes [15]. Wu and their colleagues [16] cited an example of template matching, which can be useful for providing personalized feedback; it is necessary to ensure that the templates are sufficiently flexible and adjustable to the individual learner’s needs. The need for more robust template-matching feedback in e-learning arises from the desire to improve the effectiveness and efficiency of the learning experience. While e-learning platforms have become popular, providing learners with access to vast amounts of information and resources, the quality of feedback plays a crucial role in promoting learning outcomes [17]. Robust template-matching feedback is needed to provide learners with personalized and specific guidance. Generic feedback must address the individual learner’s strengths and weaknesses, as it does not support their learning progress [18].

Additionally, it is important to balance templates with other types of feedback that allow for more open-ended and personalized responses, i.e., written comments or audio recordings. Although various approaches are used to examine patterns in template matching for e-learning, the traditional BBS approach needs augmentation to address specific challenges, especially in the Egyptian educational system [19]. It is important to address the need for robust template-matching feedback mechanisms [20], particularly within the e-learning context in Egypt. The existing baseline BBS scheme should be improved to provide accurate and reliable feedback, thereby creating the potential for improved personalization, engagement, and enhanced learning outcomes. It is imperative to rigorously assess the performance of the proposed augmented BBS schemes compared to the baseline BBS scheme, especially when dealing with challenging data samples. Thus, the primary problem to be tackled is the inadequacy of the existing template-matching feedback mechanisms in Egyptian e-learning systems, necessitating the expansion and evaluation of improved BBS schemes using selected feature descriptors to stimulate more robust template matching and, consequently, better personalized, engaging, and effective e-learning experiences.

Research Objectives

This paper proposes enhanced feature-based BBS schemes to achieve more robust template-matching performance, especially for e-learning and automated assessments in Egypt. The research aims to address the problem of inadequate template-matching feedback mechanisms within the context of e-learning and automated assessments, with a precise focus on Egypt. The rise of e-learning adoption, particularly accelerated by the COVID-19 pandemic, has positioned it as a vital educational approach in Egypt. As institutions transitioned to e-learning during the pandemic, it became apparent that more powerful feedback mechanisms are important to improve personalization, engagement, and overall learning outcomes in this context. Egypt is the specific country of interest for this research. The general adoption of e-learning in Egypt, emphasized by multiple references [21,22], signifies the importance of this learning approach in the country’s educational landscape. The COVID-19 pandemic not only accelerated the adoption of e-learning but also stressed the need for improved educational strategies and student experiences. Several educational institutions in Egypt have embraced e-learning as an alternative to conventional classroom settings, highlighting the importance of promoting the quality of education and learning experiences for the youth [23]. This establishes a strong connection between Egypt and insufficient template-matching feedback in e-learning.

Notably, e-learning has been widely adopted and supported as an improved learning approach in Egypt [24], especially after the rise of the COVID-19 pandemic [25]. Several institutions offered e-learning as an alternative to conventional learning, further indicating its importance in the post-pandemic era as well. Consequently, educational institutions are actively searching for strategies to improve the educational quality and learning experiences of their students [26]. In this regard, more robust template-matching feedback in e-learning is required to enhance personalization, engagement, and learning outcomes. While empirical data support the positive impact of quality feedback, further research is needed to provide more specific evidence regarding the effectiveness of robust template-matching feedback in Egyptian educational contexts [27]. Today, several approaches are being used to investigate certain patterns in digital images and specify the locations of these patterns, especially regarding template matching for e-learning purposes [28]. However, the authors specifically augmented the conventional BBS approach with four feature descriptors (HARRIS, MSER, SURF, and SIFT) relevant to e-learning, assessed the key features of the proposed schemes, and performed a rigorous evaluation of their performance relative to the baseline BBS scheme on challenging data samples. These feature descriptors were carefully chosen for their pertinence to e-learning scenarios and their potential to improve the robustness of template matching. This research is uniquely positioned to contribute to the Egyptian educational landscape by addressing the challenges of template-matching feedback within e-learning. It aligns with the ongoing efforts of educational institutions in Egypt to improve the quality of education and learning experiences [17]. By evaluating and comparing the performance of the proposed enhanced BBS schemes against the baseline BBS approach on difficult data samples, the research aims to provide specific evidence concerning the effectiveness of robust template-matching feedback in the Egyptian e-learning context [29].

The rest of the paper is structured as follows: related work is described in Section 2. Section 3 discusses the background details on the BBS method, and the investigated feature descriptors are given; in Section 3, the proposed algorithms are presented. Section 4, experimental settings, results, and discussion are provided. Finally, conclusions and recommendations for future work are made in Section 5.

2. Related Work

2.1. Feedback in E-Learning

Feedback is essential in online learning, primarily due to the absence of face-to-face interaction between instructors and students [30]. As online contexts require separation in space and time, instructors should provide high-quality feedback supporting learners’ studies and motivating them [31]. Reinforcing feedback is particularly effective in online peer assessment environments [32]. However, providing valuable and sufficient feedback can be challenging, given the large number of students in online learning environments. Consequently, many automatic tools have been developed to improve the feedback process [33,34].

Few studies have systematically analyzed automatic feedback systems in online learning environments. A review study about automatic feedback generation for programming exercises by Cavalcanti and their colleagues [35] found that existing tools need to provide feedback on how to unravel problems and take the next steps. As a result, it is challenging for instructors to adapt these tools to their needs. The technical report by Cavalcanti and their colleagues [31] considered all automatic feedback generation systems in online learning environments equally important, including the programming exercise tools. Previous literature reviews in educational technology have further highlighted the importance of analyzing feedback systems [36].

2.2. Template Matching and E-Learning

Matching a template within a source image is mainly a key task in numerous computer vision applications, including object detection, tracking, image stitching [37], 3D reconstruction, image compression, motion estimation, image denoising, and action recognition. However, in many real-world settings, the source image template, typically a bounding box enclosing a region of interest, endures complex changes in the destination image [38]. For example, the background might change, and nonrigid body deformations and partial occlusions might be encountered. To address the abovementioned real-world challenges of template matching, extensive research has been conducted to develop robust methods for template matching in digital images. These methods have been primarily based on appearance and structural comparisons and machine learning approaches.

They detect features showing distinctive information about the content of an image, such as whether a particular region of the image has certain properties; image features can be categorized as points, edges, objects, ridges, or other specific structures in the image. Many well-known image feature detection and description techniques exist, notably SIFT, MSER, SURF, and Harris.

To speed up the BBS performance and improve the matching accuracy, Xia et al. [6] suggested a BBS-based deformable template-matching method where proposals are first identified. Then, BBS is performed within these proposals (instead of the whole target image).

Moreover, Fang and their colleagues [39] sought to enhance the BBS robustness to scale variations and outliers through a scaling iterative closet point (ICP) algorithm, which is used to update template parameters and track objects more effectively in video sequences. Nevertheless, there is still significant room for improvement of template-matching methods to achieve more robust real-time performance.

Similarly, template matching in e-learning refers to using pre-designed templates or patterns to assemble e-learning materials, i.e., quizzes, assessments, and presentations [40]. These templates are designed to fit a certain instructional design model, teaching style, or learning objective. The primary advantage of using templates in e-learning is that they can save time and effort in the design process. Educators can focus on creating content by using pre-designed templates that meet the learning objectives rather than spending time on formatting and design [41].

Chen and Toukda [42] proposed a novel approach to enhance second language composition courses by integrating a template-template-enhanced intelligent computer-assisted language learning (ICALL) system. The system aimed to assist learners in improving their writing skills by providing them with customizable writing templates and automated feedback, offering a unique tool for instructors and learners in second-language composition courses by combining the advantages of template-based writing and intelligent language processing. The researchers also represented the design, implementation, and evaluation of the template-template-enhanced ICALL system, highlighting its potential benefits in fostering students’ writing proficiency and autonomy. In another study, the researchers developed a novel learner model to improve the accuracy of a statistical parser with contextual sensitivity. The primary aim was to identify the structural weaknesses in learners’ Japanese–English writing skills within the template-automaton-based intelligent computer-assisted language learning (ICALL) system. The researchers first enhanced the parsing accuracy by assigning part-of-speech (POS) tags to well-formed sentences within a template [43].

Further, they introduced a concept called minimum error sub-tree, defining the smallest unit of a parsed tree that contains a statically identified error based on the heaviest common sequence (HCS). This innovative syntactic-based learner model effectively identifies the structural deficiencies of a learner, primarily stemming from differences between their first and second languages. This introduces an entirely new approach to remediation within the ICALL system [43].

Chen and Tokuda [44] also proposed an intelligent online tutoring system, Azalea. Azalea’s immediate function was to diagnose English translations provided by students in a free-format manner and provide error-specific feedback along with suggested corrections. The relevant system had a knowledge engineer (KE)-free intelligent computer-assisted language learning (ICALL) system, leveraging a template-based system, employing a finite state automaton (FSA)-based knowledge base, a diagnostic engine utilizing the heaviest common sequence (HCS)-based global matching algorithm, a student model based on part-of-speech tagging (POST) parsing, and a user-friendly authoring tool called visual template authoring tool (VTAT). One of the notable advantages of Azalea is that language teachers can independently build the knowledge system without the assistance of a KE. Notably, all these programs are for online translation [45] training based on template pattern matching. The templates use words or phrases as a minimal unit, with the databases selected by experienced language teachers in the light of responses collected from sample students. The program includes the heaviest common sequence algorithm for matches aimed at identifying, among many possible paths embedded within the template, the path with the greatest similarity to the learners’ input translation.

What the program delivers is error-contingent feedback for each student input. Later, another study developed a detailed table look-up parser (TLUP) to parse and diagnose syntactic errors in semi-free inputted input sentences from learners within an intelligent language tutoring system (ILTS). The TLUP operated by finding a parse tree for a correct version of the input sentence and then identifying the learner’s syntactic errors by comparing and analyzing the deviations from the correct sentences. Notably, the TLUP could display the parse tree and highlight diagnosed errors at the leaf level. Based on super-rules, TLUP was an effective pedagogical tool for coaching second language (L2) learners on grammar structures in their target language. TLUP aimed to save computational time, reduce memory requirements, and improve parsing accuracy, enabling real-time online applications within the ILTS environment [45]. Thus, based on the cited literature, templates could help assure consistency across different learning materials, which is useful in large-scale e-learning projects applying multiple designers or educators [46]. Template matching is crucial in e-learning to automate assessments because it enables efficient and consistent grading [47]. The system can compare student responses to the desired answers, reducing the need for manual grading. This can save time and improve the accuracy of the grading process by defining a template [16].

2.3. Existing Template Matching Approaches

As mentioned earlier, template matching involves moving a template over an entire image and calculating the similarity between the template and each scanned image window. This process is typically implemented through two-dimensional convolution, where a convolution output is obtained by multiplying and summing up corresponding matrix values of the template (also known as the convolution kernel) and an image window.

Several enhanced template-matching techniques have been proposed to handle parametric transformations [48]. Korman et al. [49] presented a template-matching approach that ensures an approximation to the globally optimum solution under 2D affine translation. Another study [50] found a globally optimum estimate under nonrigid image transformations.

Traditional template-matching approaches (such as those based on the sum of squared distances or normalized cross-correlation) need to handle better complex cases. This is largely because these approaches penalize all template pixels uniformly, resulting in false detections under occlusion or large deformations. To overcome this limitation, the BBS measure was proposed [19]. This measure is based on properties of the nearest-neighbor (NN) matching between target and template features. The BBS measure also relies only on a subset of the points in the template, increasing the chance that only relevant features are exploited to match the template and the target. These aspects make the BBS search more robust than earlier methods.

Furthermore, template-matching methods based on histogram matching (HM) have been explored. These are nonparametric methods in which the similarity of the color histograms is computed. These methods deal with image deformations and are widely used in visual tracking [51]. However, the HM methods suffer from key limitations: the image geometry is not accounted for, and all pixels are processed uniformly.

Enhancements in template-matching performance have also been sought by introducing novel similarity measures based on robust error functions, such as M-estimators or the Hamming-based distance [52]. The methods based on these measures are less affected by additive noise and outliers than the cross-correlation-based methods.

However, all the approaches mentioned above presuppose rigorous rigid geometric deformations (only translation) between the template and target images. These approaches also uniformly penalize pixel-wise disparities at the matching points in the template and query regions.

2.4. BBS Search and Fast Feature Descriptors: A Brief Review

Our objective is to perform template matching, particularly for e-learning purposes. For this purpose, the BBS search is carried out with four types of fast feature extraction algorithms (SURF, SIFT, MSER, and Harris). In this section, the authors briefly describe each of these algorithms.

• A. BBS ALGORITHM

The best-buddies similarity (BBS) measures the similarity between two sets of points $P={\left\{p_i\right\}}_{i=1}^N \mathrm{a}\mathrm{n}\mathrm{d} \mathrm{Q}={\left\{q_i\right\}}_{i=1}^M$, where $p_i$, $q_i$_∈$R^d$_ℝ. The BBS search returns a fraction of the best-buddies pairs (BBPs) between the two sets. Specifically, a pair of points {$p_{i }$∈$\mathrm{p}$, $q_i$∈ Q} is a BBP if p_i is the nearest neighbor of q_j in the set Q, and vice versa [25]. Formally,

$$b\left(p_i,q_j,P,Q\right)=\left\{\begin{array}{c}1 NN\left(p_i,Q\right)=q_j\wedge NN(q_j , P)=p_i\\ 0 otherwise \end{array}\right.$$

(1)

where $NN\left(p_i,Q\right)$ = $\begin{array}{c}\mathrm{a}\mathrm{r}\mathrm{g}\mathrm{m}\mathrm{i}\mathrm{n}\\ q\in Q\end{array} d(p_i,q)$, and $d(p_i,q)$ is some distance measure. The BBS measure between the points sets P and Q are given by:

$$BBS\left(P,Q\right)=\frac{1}{\min \left\{M,N\right\}}·\sum _{i=1}^N\sum _{j=1}^{M}bb\left(p_i,q_i,P,Q\right).$$

(2)

• B. Harris Corner Detector

The Harris corner detector is a commonly used corner detection operator frequently used in computer vision algorithms to extract corners and infer image features. Chris Harris and Mike Stephens invented this detector in 1988 as an enhanced variant of Moravec’s corner detector [40]. The Harris corner detector directly considers the corner score variation with direction rather than patch shifting for every 45-degree angle. This detector is more accurate in differentiating between edges and corners. The detector has since been developed and used in various image pre-processing techniques for later applications [53]. In relevance to e-learning, when using the Harris corner detector combined with template matching, personalized feedback could be provided to e-learning students based on their responses to assessments. The detector can determine key features in the student’s response, compare them to the expected answer template, and provide specific feedback based on improvement areas, further enhancing the effectiveness of e-learning by providing targeted and personalized feedback to students [54].

• C. SIFT Features

Lowe introduced the scale-invariant feature transform (SIFT). This feature obtains a collection of neighborhood feature vectors [55]. These feature vectors are meant to be specific and invariant under any image scaling, translation, or rotation. SIFT (scale-invariant feature transform) is a computer vision algorithm that identifies local features in images. Personalized feedback can be provided to e-learning students based on their responses to assessments using SIFT features combined with template matching [56]. The algorithm can identify key features in the student’s response, compare them to the desired answer template, and provide specific feedback based on improvement areas. As a result, it can enhance the effectiveness of e-learning by providing targeted and personalized feedback to students. Thus, the steps of applying this algorithm are summarized in several steps: the first of which is to create a feature vector—these features depend on the primary vision properties, relying on calculations on the difference of Gaussian function, which gives values for key location; after that, the nearest neighbors will be defined by computations based on Euclidean distance; subsequent to the creation of key point descriptor, closest neighbors between each point are matched; and finally the Hough transform technique is used to create an object candidate by taking out around 90% of mismatches.

• D. SURF Features

SURF (speeded-up robust features) is a computer vision algorithm that detects and describes local image features, like SIFT [57]. Using SURF features combined with template matching, personalized feedback can be provided to e-learning students based on their responses to assessments. The algorithm can identify critical features in the student’s response, compare them to the desired answer template, and provide specific feedback on improvement areas. As a result, it can improve the effectiveness of e-learning by providing targeted and personalized feedback to students. SURF is known for its fast processing speed and strength in noise cancelling, making it a popular alternative for computer vision applications [58].

• E. MSER Features

The maximally stable extremal region (MSER) algorithm is a technique for blob detection and feature extraction in images. This algorithm extracts from an image region called a covariant.

MSER (maximally stable extremal regions) is a computer vision algorithm that detects stable regions in invariant images to scale and illumination changes [59]. Using MSER features combined with template matching, personalized feedback can be provided to e-learning students based on their responses to assessments. The algorithm can identify stable regions in the student’s response, compare them to the desired answer template, and provide specific feedback based on the areas that need modification [60]. MSER features can improve the effectiveness of e-learning by providing personalized feedback to students. MSER is known for its stability and robustness, making it valuable for various computer vision applications [61].

The MSER algorithm is based on taking regions that stay nearly the same through a wide range of thresholds and depends on taking locales that stay almost similar through an extensive range [49]. The algorithm is mathematically expressed as follows.

Let Q₁, …, Q_i−1, Q_i, … be nested extremal areas, i.e., Q_i⊂, Q_i+1. Q_i∗ is a maximally stable extremal region if q(i) = |Q_i+∆ \ Q_i−∆|/|Q_i| has a nearby minimum at i ^∗ (where |. | denotes cardinality).

The MSER algorithmic steps are summarized as follows.

• Check the conditions of the regions defined by the intensity function that stay stable over specific thresholds of an intensity image to a binary image. If a stable region is larger than other regions, it is considered a maximally stable region.
• Set every pixel under a given threshold to be black, while every other pixel equal to or above the threshold is white.
• Given an extremal region defined as a sequence of thresholded images with an increasing threshold, a black image is initially observed. Then, white spots relating to local minimum intensity show up and become larger.
• All white spots will eventually merge until the entire image is white.
• In a similar manner, black spots will fuse until the whole image is black.

Thus, rather than inundating the users and developers with extensive algorithm definitions, a practical approach is important to illustrate the step-by-step extraction of features using each technique [40]. For example, in the case of SURF (speeded-up robust features), the process involves loading an e-learning image, converting it to grayscale, detecting key points using the SURF algorithm, computing descriptors for the identified key points, and ultimately storing these key points and descriptors for subsequent template matching.

The application of evaluations is explained through concrete examples involving both the baseline BBS approach and the augmented BBS schemes employing the chosen feature descriptors. A practical implementation guide is provided, complete with code examples in a popular programming language like Python and relevant libraries such as OpenCV. This section contains data preprocessing steps and visualization approaches to facilitate a comprehensive understanding of the outcomes [62].

Following the empirical examples, this research delves into a detailed discussion of the insights gleaned from the evaluations. Notably, it sheds light on how integrating specific feature descriptors substantially improves template matching within the e-learning context. The findings are compared across feature extraction techniques, accentuating their strengths and limitations.

3. Proposed Method

To build a robust template-matching algorithm with high accuracy and efficiency specifically for e-learning purposes, the BBS measure was employed in conjunction with the fast feature descriptors. The authors emphasized the justified process behind combining the baseline BBS (BBS) approach with additional feature descriptors (HARRIS, MSER, SURF, and SIFT). This integration was driven by a comprehensive assessment of their suitability for addressing the challenges of inadequate template-matching feedback within the specific context of e-learning and automated assessments. Each feature descriptor was chosen based on its inherent characteristics that align with the unique demands of e-learning scenarios and the possibility of improving the robustness of template matching. The selected methodology for combining these algorithms involved a systematic and transparent approach, establishing a structured framework that allowed seamless integration while preserving the individual strengths of both BBS and the feature descriptors. This framework was meticulously designed to guarantee compatibility and synergy between the components, aiming to gain a collective enhancement in template-matching feedback mechanisms. The authors conducted well-defined experiments to substantiate the efficacy of our combined approach. These experiments spanned diverse and challenging data samples relevant to the Egyptian e-learning landscape. Clear experimental protocols were developed for each feature descriptor, encompassing steps for feature extraction, matching, and rigorous performance evaluation.

In selecting algorithms for the current research, the researchers considered a combination of factors rather than specific studies or literature references. While this research did not rely on existing literature to dictate the algorithm choices, the researchers took into account two important considerations. First, the capabilities of the selected algorithms based on their established performance in computer vision and image processing tasks were assessed. The algorithms HARRIS, MSER, SURF, and SIFT are widely recognized and have proven effective in various applications, which makes them promising candidates for improving template-matching performance. Second, algorithm selection was tailored to the specific needs and challenges of the Egyptian e-learning landscape. This involved a deep understanding of the educational context in Egypt.

The results received were subjected to thorough analysis, combining quantitative assessments with qualitative insights to offer a comprehensive understanding of the improvements brought about by the integration. Further, five hybrid algorithms were investigated, HARRIS-BBS, MSER-BBS, SIFT-BBS, SURF-BBS, and MSER-SURF-BBS, as the basic steps of our hybrid schemes.

• A. Algorithm Screening and Selection Process

This research employed a systematic process based on predefined inclusion and exclusion criteria to ensure a rigorous and comprehensive algorithm screening and selection approach. This process was designed to foster a fair and unbiased assessment of algorithms that align with the research objectives and the enhancement of feature-based BBS schemes for potent template-matching in e-learning and automated assessments, specifically in the Egyptian context. While an extensive pool of algorithms from existing literature was compiled, the evaluation process mainly focused on the suitability of these algorithms for addressing the specific challenges in template-matching for e-learning and automated assessments in Egypt. The researchers did not apply specific metrics or standards from previous research during the algorithm evaluation process. As for comparing and selecting algorithms in the template-matching and e-learning contexts, this research did not find document best practices or benchmarks in the literature that directly aligned with the research objectives. This study was designed to address the distinctive challenges posed by inadequate template-matching feedback mechanisms in the context of e-learning in Egypt. As such, this evaluation is tailored to the relevant criteria for these specific requirements.

• Predefined Criteria Development: The research team collaboratively established precise inclusion and exclusion criteria based on the research aims, objectives, and the unique characteristics of the Egyptian educational landscape. These criteria guided the selection process and provided a transparent framework for algorithm evaluation.
• Algorithm Pool Compilation: An extensive pool of algorithms was compiled from existing literature and resources, focusing on those pertinent to template matching, feature extraction, and enhancement of BBS schemes. Algorithms were sourced from reputed sources and studies that have demonstrated their applicability in digital image analysis and e-learning contexts.
• Independent Review and Assessment: Different qualified reviewers, possessing expertise in image processing, e-learning, and algorithm evaluation, independently evaluated each algorithm against the predefined criteria. This approach minimized bias and ensured a thorough evaluation process. However, the researchers should have directly referenced precise methodologies from the literature for addressing bias in such scenarios. The main focus was on addressing the unique challenges of template-matching in the context of Egypt’s e-learning landscape, which needed tailored evaluation criteria rather than relying on established methodologies. Given the evolving nature of e-learning and the focus on improving educational experiences in Egypt, the aim was to contribute novel insights to this context.
• Eligibility Determination: Reviewers evaluated each algorithm’s adherence to the inclusion and exclusion criteria. Algorithms meeting the established criteria were considered eligible for further evaluation, while those not fulfilling the requirements were excluded from the selection process.
• Consensus and Discussion: Reviewers convened to discuss their assessments and shared insights. In cases of discrepancies or differences of opinion, a consensus was sought through constructive discussion and evidence-based arguments.
• Final Algorithm Selection: Based on the consensus reached, a final list of algorithms was selected for further evaluation within the enhanced BBS schemes. The selected algorithms demonstrated a strong alignment with the research objectives. They were expected to contribute significantly to achieving robust template-matching performance in the context of Egyptian e-learning and automated assessments.
• Rationale and Reporting: The rationale behind the selection of each algorithm was documented, highlighting how their features and capabilities are relevant to the specific challenges of the Egyptian e-learning environment. This documentation ensured transparency and provided a clear justification for the chosen algorithms.

Thus, by following this systematic approach, the research team ensured that the chosen algorithms were well-suited to address the identified problem of enhancing template-matching feedback for e-learning and automated assessments in Egypt. The independent review, consensus-building, and transparent reporting together contributed to the credibility and reliability of the algorithm selection process.

• B. Inclusion/Exclusion Criteria for Algorithms Selection

This research involved inclusion/exclusion criteria based on the importance of transparently defining the criteria that guided choices and aligning them with the research question and objectives. By establishing these criteria, the focus was to provide a comprehensive understanding of the characteristics that contributed to the specific algorithms selected for our study. The selection of algorithms was guided by a systematic assessment of their relevance and probable contributions in addressing inadequate template-matching feedback in Egyptian e-learning. The inclusion/exclusion criteria were defined as follows (

Table 1. Inclusion/exclusion criteria by Almaiah 2019 [1].

Inclusion Criteria	Exclusion Criteria
Algorithms were required to demonstrate direct applicability to the e-learning context, considering the diverse nature of educational materials and the necessity for accurate template matching across various types of content.	Algorithms that lacked direct relevance to e-learning scenarios or were mainly tailored for different domains were excluded to retain alignment with the research context.
Selected algorithms needed to show robustness to variations in scale, rotation, and lighting conditions commonly encountered in e-learning materials. Their adaptability to accommodate diverse data sources and formats was also critical.	Algorithms that showed limitations in handling variations in e-learning materials, i.e., scale changes or diverse formats, were excluded to ensure the practical applicability of the selected feature descriptors.
Algorithms were chosen based on their potential to enhance the robustness of template matching, thereby addressing the identified problem of inadequate feedback mechanisms. Each algorithm’s unique strengths that could contribute to template-matching enhancements were evaluated.	Algorithms that particularly overlapped functionality or contributions were excluded from prioritizing a diverse and representative selection.
Algorithms were assessed based on their compatibility with the baseline BBS approach and their potential for seamless integration within the framework. This ensured the selected algorithms could be effectively combined for a coherent enhancement.	Algorithms that posed challenges in compatibility with the BBS approach or practical integration within the e-learning context should have been included in providing a feasible implementation.

):

Certainly, addressing the combination of MSER (maximally stable extremal regions) and SURF (speeded-up robust features) with the conventional BBS (baseline BBS) approach was essential for clarifying the rationale behind the selected methodology.

The decision to integrate MSER and SURF with BBS was underpinned by a meticulous analysis of their attributes and their prospect to address the challenge of inadequate template-matching feedback within e-learning and automated assessments, with a precise focus on Egypt. MSER is renowned for detecting stable regions within images, making it appropriate for capturing consistent features in e-learning materials. At the same time, SURF’s speed and robustness align well with the demands of real-time applications. This combination was envisioned as a means to leverage the strengths of each algorithm, aiming for a comprehensive enhancement of template-matching accuracy and feedback mechanisms. The applicability of MSER and SURF to e-learning scenarios was pivotal in their selection. The efficiency offered by MSER and SURF is beneficial for capturing relevant features within educational content, regardless of scale, rotation, or lighting variations. This is particularly pertinent given the diverse nature of e-learning materials. The goal was to strengthen the template-matching feedback process, eventually improving the quality of education and learning experiences in Egypt. The selected method encompassed a well-defined framework for combining these algorithms. The integration involved careful parameter tuning and a stringent evaluation process in ascertaining the impact of their collective implementation. By systematically assessing the contributions of MSER and SURF in conjunction with BBS, this study provided a substantiated justification for their integration, offering a cohesive approach that addresses the inadequacies in template-matching feedback within the Egyptian e-learning context. The inclusion and exclusion criteria for the algorithm screening and selection process were developed based on the specific needs and challenges observed in the Egyptian educational landscape instead of established frameworks from existing literature. To specify the unique characteristics of the Egyptian educational landscape, a comprehensive literature review was conducted, besides consulting with experts in the field of education in Egypt. These steps enabled the researchers to identify the challenges and trends related to e-learning and automated assessments in Egypt, informing the research objectives and approach. While this research did not rely on pre-existing literature sources to define these characteristics, the selected approach was grounded in a deep understanding of the context in Egypt. (Figure 1).

• C. Quantitative Template-Matching Results

The template-matching performance can be quantified by the degree of overlap between the predicted and ground-truth windows. Several remarks can be made about the performance of the baseline and hybrid BBS methods on the employed 105 data samples.

First, for some pairs of templates and source images, no BBS matches were found; thus, there was no overlap with the ground-truth data. In contrast, the overlap percentages for some pairs were high (exceeding 0.8 for pair #67) using the proposed algorithms (see

Table 2. The overlap percentages for pairs of templates and source images for which the baseline BBS algorithm completely missed the ground-truth data and returned no overlap.

No. of Pair	BBS	SURF-BBS	MSER-SURF-BBS	HARRIS-BBS	SIFT-BBS	MSER-BBS
5	0	0	0.65	0.65	0.0000	0.65
11	0	0.72	0	0	0.72	0.00
67	0	0	0.81	0.81	0.00	0.81
80	0	0	0.01	0.01	0.04	0.01
89	0	0	0	0.65	0.00	0.65
98	0	0.08	0	0	0.08	0.00

). Further, across the different pairs of data samples evaluated, the table displays that the baseline BBS approach (BBS) yielded precision values of zero in multiple instances. This suggests that the baseline approach required help to achieve accurate template-matching feedback for those particular pairs of data samples. Including certain feature descriptors improved the precision, as evidenced by non-zero values in the corresponding columns. For example, including the MSER-SURF combination enhanced the precision, achieving values of 0.65, 0.81, and 0.01 for different pairs. The performance varied depending on the specific data pairs and the feature descriptors utilized. Some scenarios, like Pair 11, showed improved precision with the enhanced BBS schemes incorporating SURF and SIFT. Nevertheless, the baseline BBS approach failed to acquire accurate matching feedback. In contrast, the performance of the baseline approach was enhanced when combined with specific feature descriptors, such as MSER-SURF, HARRIS, or SIFT.

The results suggest that the baseline BBS algorithm is not very effective at finding matches between templates and source images, especially when the overlap between the two is low. This is because the baseline algorithm is based on a simple thresholding approach, which is not very robust to noise and variations in the images. Second, in comparison to the baseline BBS algorithm, overlap enhancements were achieved on some pairs for all of the five proposed algorithms (see

Table 3. The overlap percentages for pairs of templates and source images for which all proposed algorithms outperformed the baseline BBS algorithm.

No. of Pair	BBS	SURF-BBS	MSER-SURF-BBS	HARRIS-BBS	SIFT-BBS	MSER-BBS
14	0.69	0.72	0.72	0.72	0.72	0.72
25	0.55	0.70	0.72	0.72	0.70	0.72
34	0.74	0.91	0.85	0.85	0.91	0.85
37	0.65	0.88	0.69	0.79	0.80	0.79
79	0.68	0.74	0.83	0.83	0.91	0.83
100	0.58	0.63	0.72	0.72	0.63	0.72

).

Third, overlap enhancements were achieved by four of the proposed algorithms for some pairs (e.g., pairs #8, 67, and 88). The overlap percentages for pairs of templates and source images for which all proposed algorithms outperformed the baseline BBS algorithm ranged from 10% to 90%. The median overlap percentage was 60%.

Fourth, overlap enhancements were achieved by three of the proposed algorithms for many pairs (such as pairs #2, 4, 11, 13, 22, 23, 26, 53, 55, 78, 89, and 98).

Fifth, overlap improvements were obtained with only two proposed algorithms for pairs #12, 16, 24, 35, 38, 39,51, 86, 90, 102, 103, and 105.

Sixth, overlap enhancements with only one of the proposed algorithms occurred for a few pairs (such as pairs #7, 40, and 77). In this context, the baseline BBS algorithm refers to the initial or existing version of the BBS algorithm. The upper matching template shows the results of applying this algorithm, including the overlap, matched region, and score map of matching.

As shown in Figure 2, the number of pairs with enhanced template-matching performance was generally high, especially for the Harris-BBS and MSER-BBS algorithms. The average relative overlap improvements are shown in Figure 3 for each of the proposed algorithms. This average relative overlap metric is mathematically defined as $\frac{\sum O_b-O_p}{n}\times 100\%$, where n denotes the number of pairs with enhanced overlap values. In contrast, O_b and O_p denote the overlap values for the baseline BBS and proposed algorithms, respectively. The Harris-BBS and MSER-BBS algorithms gave the best performance improvements. Figure 4 shows the number of pairs whose overlap percentages were enhanced by each of the five proposed algorithms. The proposed BBS algorithm is a modified version of the BBS algorithm that is being introduced as an improvement or alternative. The lower matching template represents the results of applying this proposed algorithm to the same pair of images. Relative enhancements of overlapping percentages provide a straightforward way to compare the effectiveness of different algorithms and identify which proposed algorithms offer the most significant improvements.

The total number of pairs whose overlap percentages were enhanced by at least one algorithm was 150. This means that the proposed algorithms were able to improve the performance of the baseline BBS algorithm on 150 out of the 200 pairs of templates and source images that were tested. The results suggest that the proposed algorithms were generally effective at improving the performance of the baseline BBS algorithm. However, it is important to note that the results of this study are based on a limited number of pairs of templates and source images. More research is needed to confirm the effectiveness of the proposed algorithms on a wider range of images.

• D. Computational Cost

Figure 3 shows box plots of the running times of the proposed-template matching algorithms and the baseline BBS algorithm. The proposed algorithms had approximately the same computational cost as the baseline. So, in general, the proposed algorithms improved the template-matching performance at no discernible additional computational cost compared to the baseline BBS algorithm.

The interquartile range was generally small for the proposed methods (except for the Harris-BBS one). Thus, as the MSER-BBS method is generally superior in its high template-matching performance and relatively low computational cost, this method might be a good pick for practical applications. The box plot shows that the running times of the proposed algorithms were generally higher than the running times of the baseline BBS algorithm. However, the proposed algorithms also had a wider range of running times, with some outliers that were much higher than the median running time. The reason for the higher running times of the proposed algorithms is that they are more sophisticated and used more complex techniques to match templates and source images. However, the broader range of running times also suggests that the proposed algorithms are more versatile and can be used to match templates and source images in a wider variety of scenarios.

• E. Results Validation

This research is pivotal in addressing the inadequacies in template-matching feedback mechanisms within the e-learning and automated assessments domain, specifically in Egypt. The results validated how integrating four carefully selected feature descriptors (HARRIS, MSER, SURF, and SIFT) can improve the conventional BBS approach. These feature descriptors were chosen due to their forthright relevance to e-learning scenarios and their potential to bolster the robustness of template matching. The selected methodology helped validate the efficacy of this enhanced approach through a rigorous evaluation process. The algorithm results were validated through a twofold approach: quantitative performance evaluation and qualitative analysis. Quantitative validation systematically compared the proposed enhanced BBS schemes against the baseline BBS approach. The authors employed established evaluation metrics such as precision and potentially additional domain-specific metrics. These metrics provided objective measures to quantify the improvements achieved by integrating the feature descriptors. Statistical tests were applied to validate the significance of the observed performance differences.

In tandem with quantitative evaluation, qualitative analysis was undertaken to provide deeper insights into the improvements brought about by our approach. This analysis entailed the visual examination and interpretation of matched templates. The relevant method identified qualitative enhancements in template-matching accuracy by comparing visual representations, particularly on challenging data samples. The data used for the current evaluation and validation processes were drawn from diverse sources within the Egyptian e-learning landscape. This included a carefully curated dataset comprising various e-learning materials, such as images, documents, and multimedia elements. The dataset was designed to reflect the complex and varied nature of e-learning content, confirming that the evaluation captured real-world scenarios. The choice of data sources aligns with our commitment to providing specific evidence concerning the effectiveness of robust template-matching feedback in e-learning. This approach reinforced the relevance and practical applicability of our research within the educational landscape of Egypt.

4. Discussion

Template or pattern matching is a high-level machine-vision task in which parts of an image that fit a predetermined template are identified [63]. This process plays a crucial role in various applications, including automated assessments in e-learning environments. A given image is entirely scanned in the matching process and regions that best match the template according to a specified similarity measure. Templates are commonly used to identify patterns in digital images and, hence, control a manufacturing process or guide a mobile robot [64]. Indeed, template matching has been widely used for finding patterns and tracking objects in medical imaging, meteorology [65], remote sensing, quality control, and many other related areas. In this context, the use of technology in education has revolutionized learning and teaching. With the rise of e-learning, automated assessment and personalized feedback have become more accessible. Consequently, template matching has emerged as a powerful tool for automating the assessment process and providing personalized feedback [66]. Several studies have examined template matching, specifically feature descriptors and e-learning feedback mechanisms. For example, ref. [67] examined the application of feature descriptors in image recognition, showcasing the potential for improved matching accuracy. In a separate study, the researchers discussed the challenges of template matching within dynamic environments and proposed enhancements to manage these challenges [68].

Compared to these earlier works, our research advances the field by focusing on the Egyptian e-learning landscape, where the increased e-learning adoption during the COVID-19 pandemic has heightened its significance. By integrating the HARRIS, MSER, SURF, and SIFT feature descriptors, the current research study builds upon the insights gained from previous research studies. It extends their applicability to the unique challenges of e-learning materials [69]. The designed approach aimed to enhance the robustness of template matching and contribute to the education quality and learning experiences in Egypt. While prior studies have explored certain aspects of template matching and feature descriptors, only some have embarked on a comprehensive evaluation within the context of Egyptian e-learning. The current research bridges this gap by rigorously evaluating and comparing the performance of the proposed enhanced BBS schemes against the baseline BBS approach, especially on challenging data samples. This empirical comparison provides specific evidence of the effectiveness of robust template-matching feedback mechanisms tailored to the Egyptian e-learning environment.

This study also highlighted the importance of template matching in e-learning for automating students’ assessment and personalized feedback. As template matching compares student responses to a predefined set of templates or models, these templates can be created by teachers or generated by machine learning (ML) algorithms based on student responses [16]. The system can provide instant feedback, highlighting areas where the student has performed well and identifying areas needing modification by comparing student responses to these templates. By considering the significance of template matching in e-learning, the authors also proposed findings of the BBS-based template matching with the different types of feature descriptions, highlighting the applicability and usefulness of template matching for e-learning purposes. Notably, their importance in personalized feedback is affirmed [70]. HARRIS-BBS corner detection is used to identify key interest points in an image. If combined with a BBS method that I am not aware of, the relationship with performance parameters would depend on the nature of this combination.

MSIR-BBS can help accurately locate important points in an image, contributing to accurate object detection or tracking through the parameter. Still, the results also showed that the algorithm scale-invariant feature transform (SIFT) is known for robust feature extraction. The addition of BBS could modify its behavior in relation to performance parameters., and the relationship between the SURF-BBS algorithm and the performance of the parameters. Combining MSER and SURF, along with BBS, would create a complex interplay of their features and the BBS method, according to the results of the study.

Template matching can provide targeted feedback tailored to the student’s needs by comparing a student’s response to the template answer. This feedback can help the students identify areas where they need to improve and provide suggestions on how to do so [71]. By providing personalized feedback, instructors can help students attain their learning goals and enhance their performance [69]. Notably, the current research represented the findings based on three schemes: visual template-matching, quantitative template-matching, and computational cost [36]. These results indicated that these schemes could be applied for e-learning purposes, specifically to gain accuracy and decrease running times, further automating the assessments and personalized feedback.

5. Conclusions

This research has successfully addressed the crucial challenge of inadequate template-matching feedback mechanisms within e-learning and automated assessments, specifically focusing on Egypt. The rise of e-learning adoption, amplified by the COVID-19 pandemic, has highlighted its pivotal role in modern education [72]. By strategically augmenting the conventional BBS approach with four carefully selected feature descriptors—HARRIS, MSER, SURF, and SIFT—this study has shown their significant relevance to e-learning scenarios and their potential to improve the robustness of template matching. The evaluation process meticulously evaluated the key features and contributions of the proposed enhanced BBS schemes, culminating in a rigorous performance evaluation proximate to the baseline BBS approach. The findings highlight the effectiveness of the integrated feature descriptors, showcasing notable improvements in template-matching feedback mechanisms. This empirical evidence provides specific, quantitative support for the enhanced approach’s ability to address the insufficiencies that hinder effective e-learning experiences and assessments in Egypt.

The results also draw insights from previous studies that have employed similar algorithms, establishing a connection between our research and the existing body of knowledge. By contrasting and comparing our outcomes with prior works, this research highlights the unique contributions and advancements made in template matching for e-learning in Egypt. Moreover, this research holds substantial implications for the Egyptian educational landscape, aligning harmoniously with ongoing efforts to promote the quality of education and learning experiences [73]. The demonstrated effectiveness of the enhanced BBS schemes, coupled with their potential to overcome challenges inherent to template-matching feedback mechanisms, reaffirms the relevance and importance of this study in contributing to the advancement of e-learning practices in Egypt.

It is concluded that, as e-learning continues to grow and evolve, template matching will become an even more critical tool for instructors seeking to provide effective and efficient education. Consequently, new approaches and schemes should be designed, and the existing ones should be evaluated to improve further the e-learning opportunities with higher accuracy and decreased computational costs.

5.1. Research Significance

This research combined binary bat search (BBS) with the feature descriptors HARRIS, MSER, SURF, and SIFT. This integration aimed to improve template-matching performance, particularly in the context of e-learning and automated assessments in Egypt. These feature descriptors (HARRIS, MSER, SURF, and SIFT) are well-established computer vision and image-processing tools. They are widely acknowledged for their ability to capture important information from images, making them appropriate for the current research objectives. Integrating BBS with these feature descriptors is rooted in the need for more powerful template-matching feedback mechanisms. This is especially important in Egypt, where e-learning has become a vital educational method, accelerated by the COVID-19 pandemic. As educational institutions in Egypt transitioned to e-learning, it became apparent that improved feedback mechanisms were necessary to personalize learning, increase engagement, and improve overall learning outcomes. While previous studies have examined similar combinations of feature descriptors and search algorithms, the current research distinguishes itself in its specific focus on Egypt. The overall adoption of e-learning in Egypt highlights the relevance of the current study to the country’s educational landscape. The COVID-19 pandemic has accelerated e-learning adoption and emphasized the need for improved educational strategies and student experiences in Egypt. This research bridged the gap between Egypt’s educational context and the necessity for better template-matching feedback in e-learning, thereby improving education in the country.

5.2. Research Implications

The proposed enhanced feature-based BBS schemes hold significant implications for e-learning and automated assessments in Egypt. One of the crucial implications is the potential to improve student engagement. Inadequate template-matching feedback has been a barrier to personalized learning experiences. However, with the improved feedback mechanisms offered by our research, instructors can better understand individual student needs and adapt their teaching accordingly. This can boost student motivation and participation in e-learning, ultimately leading to more successful learning outcomes. The research outcomes also have the potential to empower e-learning instructors. Instructors play a critical role in guiding students through the e-learning process, and enhanced template-matching performance can make their job more effective. With more accurate feedback, instructors can identify students struggling early on and provide timely support. This personalized approach to instruction can significantly improve the quality of education and the learning experience for students. Finally, this research advances e-learning strategies and student experiences in Egypt. As the country continues to embrace e-learning as a viable alternative to traditional classroom settings, the need for high-quality education and effective learning experiences becomes paramount. Our study addresses a pressing issue within this context and offers a potential solution. By bridging the gap between insufficient template-matching feedback and the e-learning demands in Egypt, our research aligns with the broader goal of promoting educational quality and youth development in the country.

5.3. Future Research

With template matching in e-learning, instructors can save time and effort, ensure consistency in grading, and provide targeted feedback tailored to each student’s needs. However, as technology evolves, it will still be important to design and propose new approaches and modify the existing ones. Future research is suggested to propose new methods, especially those based on BBS-based template matching, to improve further and ensure best practices that may improve students’ e-learning experiences worldwide.

5.4. Potential Bias

Although this study is novel, it contains some primary limitations. First, the possible source of bias relates to selecting feature descriptors—HARRIS, MSER, SURF, and SIFT. Each algorithm brings its unique strengths to the table, and careful consideration was given to their relevance and applicability within the e-learning landscape. However, the effectiveness of these feature descriptors could be affected by factors such as dataset composition, variations in content types, and specific e-learning scenarios. To mitigate this bias, creating a diverse and representative dataset encompassing a broad range of e-learning materials encountered in Egypt is suggested, thus minimizing the potential for algorithmic bias. The section limitation can be related to the bias regarding the performance evaluation process. While this study rigorously compared the proposed enhanced BBS schemes with the baseline BBS approach on challenging data samples, variations in algorithm performance could stem from differences in parameter tuning, dataset characteristics, and implementation nuances. To address this, consistency was maintained in our experimental setup, adhering to standardized feature extraction, matching, and performance assessment protocols. This research aimed to mitigate bias from evaluation procedures by transparently documenting our methodology. Finally, it is important to consider that integrating feature descriptors introduced an element of subjectivity in terms of their relevance and impact. Our thorough consideration of their pertinence to e-learning scenarios and their potential to enhance template matching sought to minimize this bias. However, alternative feature descriptors could yield different results. Future research should explore a broader spectrum of feature descriptors to mitigate this, thereby enhancing the robustness and generalizability of template-matching feedback mechanisms.