**1. Introduction**

Image segmentation has potential to reduce the image complexities associated with processing of singular image primitives. Low-level segmentation of an image in nonoverlapping set of regions called super-pixels helps in pre-processing and speeding up further high-level computational tasks related to visual images. The coherence feature of super-pixels allows faster architectural functionalities of many visual applications including object localization [1], tracking [2], posture estimation [3], recognition [4,5], semantic segmentation [6], instance segmentation [7], and segmentation of medical imagery [8,9]. These applications will be aided by super-pixels in terms of boosted performances, as the super-pixels put forward only the discriminating visual information [10].

**Citation:** Hashmani, M.A.; Memon, M.M.; Raza, K.; Adil, S.H.; Rizvi, S.S.; Umair, M. Content-Aware SLIC Super-Pixels for Semi-Dark Images (SLIC++). *Sensors* **2022**, *22*, 906. https://doi.org/10.3390/s22030906

Academic Editor: Jing Tian

Received: 27 September 2021 Accepted: 17 November 2021 Published: 25 January 2022

**Publisher's Note:** MDPI stays neutral with regard to jurisdictional claims in published maps and institutional affiliations.

**Copyright:** © 2022 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (https:// creativecommons.org/licenses/by/ 4.0/).

Low-level segmentation tends to result in incorrect segmentations if the visual image has high pixel intensities; these high pixeled values are usually the biproduct of visual scenes captured in low lighting conditions, i.e., semi-dark images and dark images. The obtained incorrect super-pixels are attributed to the underlying approach used for the final segmen<sup>t</sup> creation. The existing super-pixel methods fail due to incorrect pixel manipulations for base operational functionality. The currently employed pixel manipulation relies on straight line differences for super-pixel creation. These straight-line difference manipulations fail to retain the content-aware information of the image. The results are further degraded if the image has low contrasted values, which result in no clear discrimination among the objects present. The existing super-pixel creation methods are divided into two categories based on the implemented workflow. The two categories are graph-based and gradient ascent based [10,11]. The former, focuses on minimization of cost function by grouping and treating each pixel of the image as a graph node. The later, iteratively processes each image pixel using clustering techniques until convergence [12]. One of the typical features observed for identification of super-pixel accuracy is regularity, i.e., to what extent the super-pixel is close to the actual object boundary of the image. All the graphbased methods for super-pixel calculations suffer from poorly adhered super-pixels which result in irregularity of segments presented by super-pixels [11]. Additionally, graph-based methods are constrained by the excessive computational complexity and other initialization parameters. Whereas, the gradient-ascent methods are simplistic in nature and are recommended in the literature due to resultant high performance and accuracy [13,14]. However, there are some issues associated with content irrelevant manipulation of singular pixels to form resultant super-pixels. Some of the key features of using super-pixel segmentations are:


With these advantages of super-pixels, they are highly preferred. However, superpixel abstraction methods backed by gradient-ascent workflows are also limited in their working functionality to retain the contextual information of the given image [15]. The contextual information retainment is required to achieve the richer details of the visual image. This loss of contextual information is caused by flawed pixel clusters created based on Euclidean distance measure [16]. As the Euclidean distance measures calculates straight line differences among pixels which ends up in irregular and lousy super-pixels. Moreover, further degradation can be expected to process the semi-dark images where high pixel intensities along with no clear boundaries are observed. In such scenarios, the propagation of inaccurate super-pixels will affect the overall functionality of automated solutions [11]. To overcome this problem of information loss for creation of compact and uniform superpixels, we propose content-aware distance measures for image pixel cluster creation. The content-aware distance measure as the core foundational component of gradient-ascent methods for super-pixel creation will not only help in alleviation of information loss, but it will also help in preserving less observant/perceptually visible information of semi-dark images. The state-of-the-art methods for super-pixels creation have not been analyzed exclusively on the semi-dark images which further raises the concerns related to segmentation accuracy. In nutshell, the problems in existing segmentation algorithms are:


To resolve these issues, the presented research exclusively presents multifaceted study offering following features:

1. Classification of Literary Studies w.r.t Singular Pixel Manipulation Strategies:

The categorization of the existing studies is based on entire image taken as one entity. The image entity can represent either graph or a feature space to be clustered, i.e., graph-based or gradient-ascent based methodology for pixel grouping. To the best of our knowledge, there has been no study that categorizes existing studies based on the manipulation strategy performed over each pixel. We present the detailed comparative analysis of existing research highlighting their core functionality as the basis for classification.

2. Investigation of Appropriate Pixel Grouping Scheme:

The grouping scheme backing the image segmentation module being crucial component can highly affect the accuracy of the entire model. For this reason, to propose a novel extension as a generalized solution of all types of images including semi-dark images, we present a detailed qualitative investigation of up to seven distance measures for grouping pixels to create super-pixels. The investigation resulted in shortlisted pixel grouping measures to retain fine grained details of the visual image.

3. Novel Hybrid Content-Aware Extension of SLIC—SLIC++:

SLIC, being the simplest and fastest solution for the pixel grouping, remains the inspiration, and we enhance the performance of SLIC by adding content-aware feature in its discourse. The proposed extension holds the fundamental functionality with improved features to preserve content-aware information. The enhancement results in better segmentation accuracy by extracting regular and continuous super-pixels for all scenarios including semi-dark scenarios.

4. Comprehensive Perceptual Results focusing Semi-dark Images:

To assess the performance of the proposed extension SLIC++ for extraction of the richer information of the visual scene, we conduct experiments over semi-dark images. The experimental analysis is benchmarked against the standard super-pixel creation methods to verify that incorporating content-aware hybrid distance measure leads to improved performance. The perceptual results further conform better performance, the scalability, and generalizability of results produced by SLIC++.
