ESFD-YOLOv8n: Early Smoke and Fire Detection Method Based on an Improved YOLOv8n Model

Abstract

Ensuring fire safety is essential to protect life and property, but modern infrastructure and complex settings require advanced fire detection methods. Traditional object detection systems, often reliant on manual feature extraction, may fall short, and while deep learning approaches are powerful, they can be computationally intensive, especially for real-time applications. This paper proposes a novel smoke and fire detection method based on the YOLOv8n model with several key architectural modifications. The standard Complete-IoU (CIoU) box loss function is replaced with the more robust Wise-IoU version 3 (WIoUv3), enhancing predictions through its attention mechanism and dynamic focusing. The model is streamlined by replacing the C2f module with a residual block, enabling targeted feature extraction, accelerating training and inference, and reducing overfitting. Integrating generalized efficient layer aggregation network (GELAN) blocks with C2f modules in the neck of the YOLOv8n model further enhances smoke and fire detection, optimizing gradient paths for efficient learning and high performance. Transfer learning is also applied to enhance robustness. Experiments confirmed the excellent performance of ESFD-YOLOv8n, outperforming the original YOLOv8n by 2%, 2.3%, and 2.7%, with a mean average precision ([email protected]) of 79.4%, precision of 80.1%, and recall of 72.7%. Despite its increased complexity, the model outperforms several state-of-the-art algorithms and meets the requirements for real-time fire and smoke detection.

1. Introduction

Despite fire being an essential tool for human development, providing warmth and light, and enabling various technological advancements, it can quickly transform into a destructive force when it escapes our control. Recent statistics underscore the devastating impact of uncontrolled fires on both human lives and infrastructure. In 2022, the United States experienced an estimated 1.5 million fires, resulting in thousands of fatalities, injuries, and USD billions in property damage [1]. The World Health Organization reported over 180,000 annual deaths globally caused by burns and fire-related incidents, many of which occur in densely populated areas [2]. These alarming figures highlight the critical need for advanced fire detection systems capable of the early and accurate identification of fire incidents, potentially saving countless lives and minimizing property loss.

Implementing advanced smoke and fire detection systems significantly improves safety and resource management by accelerating response times and minimizing false alarms. Fire detection has long relied on traditional sensor-based methods, including heat, smoke, flame, and gas sensors. These systems, such as those described in [3,4,5,6,7], are used widely because of their effectiveness in various environments. On the other hand, these sensors often face significant drawbacks, particularly regarding response time and false alarm rates [8,9,10,11,12,13,14]. For example, heat sensors may react too slowly in fast-spreading fires, while smoke sensors can be triggered by non-fire-related particulates, leading to frequent false alarms. Moreover, flame detectors, which rely on the optical spectrum, can be ineffective in smoky conditions where visibility is low. Gas sensors, however, require specific combustible gases to be present and may not detect all types of fires efficiently.

To overcome these limitations, researchers have explored machine learning (ML) methods that analyze color, texture, and motion in visual data for fire detection. Previous studies [15,16,17,18] used color-based models to identify fire pixels, leveraging the characteristic color features of flames. On the other hand, these models often struggle with varying lighting conditions and similar-colored objects. Texture-based approaches examine the spatial patterns of fire but can be confounded by complex background textures, as discussed in [19,20]. Motion-based detection, which has been used in previous studies, capitalizes on the dynamic nature of fire, distinguishing it from static objects [21,22,23]. Nevertheless, these methods are challenged by other moving objects, such as humans or vehicles, which can result in false positives. Despite these advances, ML methods still require robust preprocessing and feature extraction, which can be computationally intensive and less effective in real-time applications [24,25,26].

The advent of deep learning (DL) has brought significant advances in fire detection technology, addressing many of the challenges faced by traditional and ML methods. In particular, convolutional neural networks (CNNs) have shown remarkable success because of their ability to extract hierarchical features automatically from raw input data. For example, the study by Saydirasulovich et al. [27] achieved an average precision (AP) of 79.4% in identifying fires and smoke within intricate scenarios. In addition, the FireNet architecture [28] is lightweight enough to be deployable on embedded platforms, such as Raspberry Pi, balancing accuracy and resource efficiency.

Efforts to boost the speed and real-time performance of fire detection algorithms have yielded promising results. Adaptations of YOLOv5 [29,30,31,32,33] and the MobileNetV2-SSD model [34,35,36] enable real-time fire detection on embedded systems and video surveillance applications, achieving a harmony between speed and accuracy. Furthermore, recent research has focused on developing fire detection systems capable of operating in diverse environments. A previous study [37] proposed a hybrid system combining local binary pattern convolutional neural network (LBP-CNN) and YOLOv5 architectures. They reported a fire and smoke detection precision rate of 96.25% for a normal scenario, 93.2% for a foggy environment, and a combined detection average precision rate of 94.59% across varied settings. This approach highlights the importance of using spatial information for accurate detection in diverse environments. The authors of [38] incorporated Ghost modules [39] and attention mechanisms, resulting in a 27% parameter reduction, a 2.9% increase in mAP50small, and an FPS of 24.4. In addition, incorporating temporal information alongside spatial data has shown promising results. Studies, such as de Venâncio [40], reported how combining both aspects helps to reduce false positives without compromising accuracy or processing time, proving advantageous over common approaches. These novel methodologies pave the way for more robust and adaptable fire detection systems.

Similarly, research has delved into wildfire detection across various fire and smoke images, tailoring deep learning architectures to differing lighting and weather conditions [41,42,43,44,45,46,47]. Other investigations have explored the use of transformer-based models, which offer improved feature extraction and classification capabilities for complex fire and smoke patterns [48,49,50]. Additionally, advancements have been made in leveraging multi-spectral and hyperspectral imaging combined with machine learning techniques to enhance the detection accuracy and responsiveness of wildfire detection systems [51,52,53].

Dedicated efforts have also been directed towards smoke detection, a fundamental piece of early fire warning systems. By incorporating techniques, such as the channel prior dilatation attention module and mixed-classification detection head, advanced object detection architectures inspired by the YOLOv8 model have demonstrated promising results in enhancing smoke detection capabilities for static images and video sequences [54]. Furthermore, the application of generative adversarial networks (GANs) has shown potential in augmenting training datasets, thus improving the robustness of smoke detection algorithms in varied conditions [55,56,57,58]. Research has also investigated the role of edge computing in smoke detection, emphasizing the benefits of distributed processing and real-time analytics in minimizing detection latency and improving system scalability [59,60,61,62].

Although AI-driven smoke and fire detection has seen significant advancements, achieving a balance between high accuracy and resource efficiency remains challenging. This paper addresses this issue by introducing a novel early smoke and fire detection system based on YOLOv8n [63], which was designed for broad applicability across diverse environments. The unique advantages of the system are highlighted through a comparative analysis with existing fire detection models, including those within the well-established YOLO family [64], all trained on the same dataset.

The contributions of this study include the following:

• The backbone network was enhanced by substituting C2f blocks with residual blocks [65]. This architectural change improves feature extraction and information flow, producing more robust representations of smoke and fire patterns;
• C2f-integrated generalized efficient layer aggregation network (GELAN [66]) blocks (C2fGELAN) were introduced in the neck network, replacing standard C2f blocks. C2fGELAN is a powerful and state-of-the-art feature enhancement block that refines feature maps, enabling the model to better distinguish among smoke, fire, and background elements;
• The standard Complete Intersection over Union (CIoU [67]) loss function was replaced with the Wise Intersection over Union version 3 (WIoUv3) [68] loss during training. The accuracy of the bounding boxes delineating smoke and fire was significantly improved, resulting in a more precise localization of these hazards;
• Transfer learning was applied, further enhancing the model’s robustness and ability to generalize across various environments and conditions;
• Extensive experiments, including comparisons with the YOLO series and other methods, showed that this model achieved the best overall accuracy and efficiency in detecting smoke and fire in diverse scenarios.

The remainder of this work is structured as follows. Section 2 details the materials and methods, including a review of relevant literature and the model architecture. Section 3 outlines the datasets, evaluation metrics, and experimental setup, followed by a presentation and analysis of the results, highlighting the strengths and weaknesses of this approach. Section 4 concludes this paper by summarizing the key findings, discussing the limitations identified through error analysis, and suggesting potential directions for future research.

2. Materials and Methods

2.1. YOLO Algorithm

The “you only look once” (YOLO) series revolutionized real-time object detection with speed, accuracy, and adaptability. Since 2015, each iteration has improved upon its predecessors. YOLO models allow real-time processing using a single-stage architecture. Despite prioritizing speed, they show competitive accuracy.

The YOLO series has seen significant advances over the years. YOLOv1 introduced the groundbreaking concept of real-time object detection with a single-stage network. YOLOv2, also known as YOLO 9000 [69], showed improved accuracy and speed through techniques such as anchor boxes, multi-scale training, and better backbone networks. YOLOv3 [70] introduced a feature pyramid network (FPN) [71] for the better detection of objects on different scales, further improving accuracy. YOLOv4 [72] incorporated various architectural and training optimizations, increasing speed and accuracy. YOLOv5 [73] focused on ease of use, model size optimization, and training flexibility, making it popular for deployment on diverse platforms. YOLOv6 [74] emerged as an anchor-free version, streamlining the detection process. YOLOv7 [75] focused on performance optimization, incorporating new training, and extending the efficient layer aggregation network (ELAN [76]) to E-ELAN, thereby improving efficiency and prioritizing both accuracy and speed. YOLOv8 integrated advances in anchor-free detection, loss functions, and training strategies. YOLOv9 introduced the novel concepts of programmable gradient information (PGI) and the GELAN, achieving state-of-the-art results on the MS COCO dataset. The most recent iteration, YOLOv10 [77], developed by researchers at Tsinghua University, eliminated the non-maximum suppression (NMS) step, optimized the architecture for enhanced performance, and reduced computational overhead in end-to-end real-time object detection.

Since its inception, researchers have continuously enhanced the accuracy and robustness of YOLO through various approaches, such as refined architectures, attention mechanisms, and optimized training strategies [78,79,80,81]. The adaptability of YOLO has been demonstrated in diverse applications such as aerial imagery [82,83,84,85], underwater environments [86,87,88], agriculture [89,90], and medical imaging [91,92,93]. Efforts to boost the speed and efficiency of YOLO for real-time use cases have involved techniques like model compression and hardware acceleration [94,95,96]. Moreover, the scope of YOLO has expanded beyond object detection, integrating with tasks like semantic segmentation and action recognition for a more comprehensive scene understanding [97,98,99,100].

2.2. YOLOv8n Network

This study focused on YOLOv8n, the nano version of the YOLOv8 object detection model developed by Ultralytics LLC located in Los Angeles, CA, United States. Despite advances in YOLOv9 and YOLOv10, YOLOv8n was selected for its proven effectiveness and balanced approach within the YOLO family. The introduction of PGI by YOLOv9 complicates training, while the elimination of Non-Maximum Suppression (NMS) by YOLOv10 raises resource requirements during training. Consequently, YOLOv8n is more suitable for resource-constrained environments while delivering competitive performance. The architecture of YOLOv8n (Figure 1) consists of the following three main components: a backbone network, a neck network, and a head network.

YOLOv8n uses a modified version of the CSPDarknet [101] architecture as its backbone. It comprises stacked modules, including CBS (light blue colored blocks in Figure 1) (convolutional layer (Conv) + batch normalization (BatchNorm [102]) + Sigmoid Linear Unit (SiLU [103])) and C2f modules (warm peachy colored blocks in Figure 1), culminating in the spatial pyramid pooling fast (SPPF) module (gray colored blocks in Figure 1). While CBS and C2f modules aid in feature extraction, the SPPF module enhances feature expression using max pooling on previously pooled features, improving module efficiency.

The neck of YOLOv8n features a combination of path aggregation network (PANet [104]) and FPN layers (Figure 2). This hybrid approach enhances the ability of the model to aggregate features from different scales, which is crucial for detecting objects of varying sizes. PANet improves information flow among layers by adding bottom-up paths, enriching the semantic information at each level. FPN, on the other hand, helps build high-level feature maps with strong semantics. Additionally, the Upsample blocks (light coral colored blocks in Figure 1) enhance feature map resolution to maintain spatial accuracy, while the Concatenation blocks (gold colored blocks in Figure 1) merge features from different layers to integrate both high-resolution and semantically rich information. Together, these components ensure that the features used by the head are rich and diverse, enhancing detection accuracy.

The head of YOLOv8n (charcoal blue colored blocks in Figure 1) predicts the bounding boxes, objectness scores, and class probabilities. It features a fully convolutional structure with multiple output layers, each corresponding to a different scale. This multi-scale prediction approach ensures the model can detect small, medium, and large objects. YOLOv8n uses an anchor-free approach that simplifies the design and reduces computational complexity. This approach predicts the bounding boxes directly without predefined anchor boxes, improving flexibility and performance.

2.3. Generalized Efficient Layer Aggregation Network (GELAN)

The efficient layer aggregation network (ELAN [76]) has revolutionized object detection tasks. ELAN enhances feature learning and computational efficiency by optimizing gradient paths within deep convolutional networks. Its integration of cross-stage partial connections (CSPs) and stacked computational blocks facilitates improved gradient flow during training, resulting in faster model convergence and excellent performance in identifying smoke and fire in their early stages. Moreover, the design of ELAN mitigates the risk of overfitting, making it adaptable to diverse datasets encompassing various smoke and fire scenarios.

Nevertheless, the focus of ELAN on stacking convolutional layers restricts the flexibility to integrate other computational blocks that might be advantageous for detecting the subtle early signs of smoke and fire. This limitation can hinder its effectiveness in complex real-world scenarios, where varying environmental conditions and potential obstructions can impede accurate early detection.

In YOLOv9, GELAN addresses these limitations (Figure 3). Building upon the strengths of ELAN, GELAN broadens the concept of layer aggregation, enabling the incorporation of any computational block, not just convolutional layers. This enhanced flexibility allows the integration of specialized modules like attention mechanisms or unique feature extractors, which are better equipped to capture the distinctive characteristics of early smoke and fire. GELAN retains the emphasis of ELAN on gradient path optimization, ensuring efficient learning and improved performance.

GELAN offers a more versatile and powerful architecture for object detection tasks because it merges the best aspects of CSPNet and ELAN. Its ability to integrate diverse computational blocks seamlessly while maintaining high efficiency and accuracy makes it an invaluable asset in pursuing enhanced early detection capabilities for critical real-time applications.

2.4. Wise Intersection over Union Version 3 (WIoUv3)

In object detection, selecting the right loss function for bounding box regression is crucial for accuracy. The primary goal is to minimize differences between the predicted boxes of the model and the actual ground truth boxes around objects. This choice is particularly important in challenging applications, such as smoke and fire detection, where training data annotations may be noisy or incomplete.

This study addressed this by replacing the standard CIoU loss function in YOLOv8n with WIoUv3. WIoUv3 offers significant advantages, particularly when dealing with suboptimal training data.

WIoUv3 features a two-layer attention mechanism and a dynamic non-monotonic focusing mechanism. The attention mechanism helps the model focus on relevant features within bounding box predictions. The dynamic focusing mechanism evaluates the quality of candidate boxes (anchors) through outlier analysis, reducing the impact of noisy annotations and preventing the model from being penalized for minor geometric discrepancies. When the predictions are close to the ground truth, WIoUv3 de-emphasizes geometric factors, aiding efficient learning and better generalization. The precise mathematical formulation of this loss function is expressed as Equations (1)–(7).

$$S_u=wh+w_{gt}{h}_{gt}-W_i{H}_i,$$

(1)

$$IoU={\displaystyle \frac{W_i{H}_i}{S_u}},$$

(2)

$$L_{IoU}=1-IoU,$$

(3)

$$R_{WIoU}=\mathrm{e}\mathrm{x}\mathrm{p}\left({\displaystyle \frac{{\left(x-x_{gt}\right)}^2+{(y-y_{gt})}^2}{W_g^2+H_g^2}}\right),$$

(4)

$$L_{WIoUv1}=R_{wIoU}{L}_{IoU},$$

(5)

$$L_{WIoUv3}={rL}_{wIoUv1}$$

(6)

$$r={\displaystyle \frac{\beta }{\delta {\alpha }^{\beta -\delta }}}$$

(7)

where $IoU$ is the intersection over union [105]; $r$ is the gradient gain, $\alpha$ and $\beta$ are the anomaly levels of the predicted box, with lower values indicating higher quality. Smaller gradient gains are assigned to boxes with larger anomalies to reduce the harmful gradients during training. $\alpha$ and $\delta$ are the hyperparameters controlling gradient gain. Figure 4 presents the other parameters, with blue and green boxes representing predicted and ground truth boxes, respectively. Coordinates are denoted by $(x,y)$ and $(x_{gt},y_{gt})$, while the heights and widths are $(h,w)$ and $(h_{gt},w_{gt})$. $H_i$ and $W_i$ are the height and width of the intersection rectangle, while $H_g$ and $W_g$ are the height and width differences between the combined boxes and their intersection. $S_u$ in Equation (1) is the union area of the predicted box and the ground truth box.

2.5. Residual Block (ResBlock)

Residual blocks (ResBlocks) are integral to deep neural networks, particularly in architectures like ResNet (Figure 5). They incorporate skip connections, enabling them to learn residual functions relative to the layer input rather than attempting to learn functions from scratch.

Therefore, residual blocks aim to fit a residual mapping of the desired underlying mapping. This is achieved by recasting the original mapping into a residual form, denoted as Equation (12).

$$f_1={CBS}_{1\times 1}\left(I\right),$$

(8)

$$f_2={CBS}_{3\times 3}\left(f_1\right),$$

(9)

$$f_3={CB}_{1\times 1}\left(f_2\right),$$

(10)

$$f_4={CB}_{1\times 1}\left(I\right),$$

(11)

$$Y=\sigma (f_3+f_4)$$

(12)

where $I$ represents the input to the block, the $CBS$ block (light blue colored blocks in Figure 5) denotes a sequence of a convolutional layer, batch normalization layer, and SiLU activation function, while the $CB$ block (powder blue colored blocks in Figure 5) consists of a convolutional layer followed by a batch normalization layer, and $\sigma$ represents the SiLU activation function.

The residual blocks comprise convolutional layers, batch normalization layers, and SiLU activation functions. SiLU is the activation function within these blocks. Batch normalization plays a crucial role in the proper initialization of neural networks, ensuring that activations throughout the network conform to a unit Gaussian distribution at the outset of training.

The rationale behind using residual blocks lies in the ease of optimizing the residual mapping compared with the original unreferenced mapping. Adjusting the residual towards zero is simpler if an identity mapping is optimal rather than attempting to learn the identity mapping directly through a stack of nonlinear layers. The inclusion of skip connections facilitates the learning of identity-like mappings, enhancing the capacity of the network to learn complex functions. Consequently, the residual acts as a correction to the original mapping.

2.6. ESFD-YOLOv8n Architecture

This paper introduces ESFD-YOLOv8n (Early Smoke Fire Detection-YOLOv8n), a specialized YOLOv8n model for early and accurate smoke and fire detection (Figure 6). The proposed model refines the following three critical components: feature extraction, the neck network, and box prediction (

Table 1. Comparison of the proposed and YOLOv8n architectures.

Features	The Proposed Architecture	YOLOv8n
Backbone	ResBlock	C2f
Neck	C2fGELAN	C2f
Bounding box loss function	WIoUv3	CIoU
Techniques used	ResBlock, WIoUv3	-
Novel technique introduced	C2fGELAN	-

). Residual blocks (ResBlock—warm peachy colored blocks in the Backbone in Figure 6) are integrated into the architecture to strengthen the ability of the model to extract meaningful features from images, making it more resilient to variations in lighting, perspective, and other visual complexities often found in challenging environments. This enhancement significantly improves the ability of the model to identify subtle visual cues associated with smoke and fire, even in their earliest stages.

In the neck network, this study introduced C2fGELAN—the integration of C2f and GELAN blocks (warm peachy colored blocks in the neck in Figure 6)—enriching extracted features with deeper semantic information (Figure 7). This comprehensive understanding of image content empowers the model to distinguish smoke and fire from other visual elements more precisely, minimizing false alarms and enhancing overall detection accuracy. The WIoUv3 loss function (light steel blue colored block in the neck in Figure 6) was used as the box predictor loss function, replacing the standard CIoU. This strategic choice significantly enhanced the accuracy of the model in detecting smoke and fire at the earliest stages. The WIoUv3 loss function provided a more refined and precise approach, allowing the model to identify potential fire hazards more accurately.

3. Experiments and Analysis

3.1. Dataset

The D-Fire dataset [106] was used to train the YOLOv8n-based model. D-Fire is a freely available community resource specifically curated for fire and smoke classification and detection tasks. The strength of the dataset lies in its diversity, containing 21,527 images encompassing a broad spectrum of scenarios. These include indoor and outdoor environments, fires of varying sizes, scenes with varying light conditions, and, importantly, normal scenes devoid of fire or smoke, which often cause false alarms in conventional algorithms. In particular, the dataset also includes images featuring objects or environments humans commonly misinterpret as fire or smoke, strengthening the ability of the model to navigate the complexities of smoke or fire detection. Figure 8 shows the balanced representation of fire/smoke and normal scenes in D-Fire, which facilitates training a model that can effectively generalize to real-world applications.

Table 2. D-Fire dataset description.

Category	Images
Fire only	1164
Smoke only	5867
Fire and smoke	4658
Background	9838

provides a detailed description of the dataset.

The D-Fire dataset was divided into 70% training, 20% validation, and 10% testing sets for our early smoke and fire detection research. This ensured sufficient training data, hyperparameter tuning with the validation set, and a separate testing set for evaluating real-world performance.

Table 3. Details of the dataset.

Sets	Images	Percentage
Training	15,069	70
Validating	4305	20
Testing	2153	10

lists the image distribution.

3.2. Evaluation Metrics

Performance metrics [107,108] are essential for evaluating the effectiveness of object detection algorithms like YOLOv8. These metrics provide insights into how well the model detects and localizes objects within images. Precision describes the percentage of detected objects that are genuinely the desired class (Equation (13)), while recall indicates the percentage of actual objects the model detects (Equation (14)). An ideal model would have both precision and recall approaching 100%.

$$Precision={\displaystyle \frac{True Positive}{True Positive+False Positive}},$$

(13)

$$Recall={\displaystyle \frac{True Positive}{True Positive+False Negative}}$$

(14)

These metrics often trade off against each other. A model may achieve high precision by making few predictions, but it might miss many objects, resulting in low recall. Average precision ($AP$) is used to balance this trade-off (Equation (15)). $AP$ summarizes the precision–recall trade-off of the model across different confidence thresholds.

$$AP={\int }_0^{1}P\left(R\right)dR$$

(15)

The mean average precision ($mAP$) provides a single number summarizing performance across all object classes (Equation (16)). The $mAP$ is the mean of $AP$ values for each class. The $mAP$ at a 50% $IoU$ threshold ($mAP@5$0) is a specific instance of the $mAP$ where the $IoU$ threshold is set to 0.5. This means a prediction is considered correct if it overlaps with the ground truth by at least 50%.

$$mAP={\displaystyle \frac{1}{N}}{\sum }_{i=1}^N{\int }_0^1{P}_1\left(R\right)dR$$

(16)

The inference time measures how long the model takes to process an image and generate predictions (Equation (17)). This metric is crucial for real-time applications, directly affecting the responsiveness of the system.

$$InferenceTime={\displaystyle \frac{TotalProcessingTime}{NumberofImages}}$$

(17)

3.3. Configuration Parameters

The proposed method in Anaconda 24.1.2 was constructed and examined using Python distribution version 3.11.7 on a computer with a 12th Gen Intel(R) Core (TM) i9-12900K (24 CPUs), ~3.2 GHz, manufactured by Intel Corporation, located in Santa Clara, United States, 64 GB RAM, and NVIDIA^® GeForce RTX™ 4060 Ti 8 GB graphics card, manufactured by NVIDIA Corporation, located in Santa Clara, United States. The model was trained with a batch size of 32 for 300 epochs, starting with an initial learning rate of 0.01.

Table 4. Model training configuration parameters.

Parameters	Values
Epochs	300
Batch size	32
Image size	640
Learning rate	0.01
Patience	50
Pretrained	yolov8n.pt

lists the configuration parameters for the model.

3.4. Ablation Study

The contribution of each proposed improvement strategy was evaluated by conducting a series of ablation studies using the D-Fire dataset on our baseline model. The results (

Table 5. Detection results after introducing different improvement strategies (the bold data in the table indicate the best results and the check mark (✓) indicates that the method in the corresponding column is applied in the corresponding row).

Baseline	WIoUv3	Residual Block	C2f GELAN	Transfer Learning	P (%)	R (%)	[email protected] (%)	Time (ms)	FLOPs (B)
YOLOv8n	-	-	-	-	77.8	70.0	77.4	0.8	8.1
	✓	-	-	-	79.1	70.6	78.0	0.8	8.1
	✓	✓	-	-	80.8	71.0	78.8	0.9	10.6
	✓	✓	✓	-	81.2	71.2	79.3	1.0	10.8
	✓	✓	✓	✓	80.1	72.7	79.4	1.0	10.8

) provide valuable insights into the efficacy of each enhancement.

Table 5 lists the incremental improvements in detection performance achieved by each strategy. In particular, integrating WIoUv3 into the prediction box regression loss enhanced localization significantly, contributing to a 0.6% improvement in [email protected]. Replacing the C2f module with residual blocks in the backbone network further refined the attention of the model to crucial features, boosting [email protected] by an additional 0.8%. The substitution of C2f layers with the C2fGELAN module in the neck resulted in a 0.5% improvement in [email protected], demonstrating the effectiveness of this module in balancing speed and accuracy while improving feature fusion. Finally, applying transfer learning led to a further 0.1% gain in [email protected], highlighting the potential of this technique for model optimization.

The proposed network boosted the detection performance, with a 2.0% average increase in accuracy and notable gains across other metrics, showing that the slight increase in the model size and parameters is a worthwhile trade-off for the substantial accuracy improvement. Furthermore, the model demonstrates proficiency in detecting small-scale objects (Figure 9).

3.5. Feature Map Visualization

The feature extraction capabilities of the baseline model were enhanced by replacing the existing C2f modules in the backbone with residual blocks and further refining the neck network by substituting C2f layers with C2fGELAN blocks. The impact of these changes was visually assessed using the Eigen-CAM [109] method, focusing on the 8th layer (backbone) and the 21st layer (neck).

Figure 10 provides a clear visualization of the improvements by comparing the proposed model and the original YOLOv8n model on the same images. The bright red and yellow areas on the feature map indicate regions where both models detected significant features, while the blue areas show regions of lower attention. The input images (Figure 10a–c) are presented with enhanced smoke and fire visibility for clarity. The differences between the C2f and ResBlock modules are shown in Figure 10d–f and Figure 10g–i, respectively. Figure 10g–i show the superior ability of the ResBlock module to capture finer details, guiding the backbone network toward more pertinent information. The original and modified necks of YOLOv8n and ESFD-YOLOv8n are depicted in Figure 10j–l and Figure 10m–o, respectively. The enhanced feature maps in Figure 10m–o show the effectiveness of the C2fGELAN upgrade, resulting in heightened resolution and facilitating efficient context aggregation and fusion within the network.

These visual results confirm that the ESFD-YOLOv8n model outperformed the original model in terms of accuracy, robustness, and efficiency.

3.6. Comparison of the ESFD-YOLOv8n with Other Detection Algorithms

The effectiveness of the proposed ESFD-YOLOv8n model was examined through an extensive evaluation of the D-Fire dataset, comparing its performance against a diverse selection of state-of-the-art smoke and fire detection approaches. These included six YOLO series variations (YOLOv5n, YOLOv5s, YOLOX tiny [110], YOLOv6n, YOLOv7 tiny, YOLOv8n, GELAN, YOLOv9, and YOLOv10n) and six other notable approaches [111,112,113,114,115,116]. The comprehensive results, highlighted in

Table 6. Comparison of the different models (Bold data in the table indicate the best results).

Methods	Precision (%)	Recall (%)	[email protected] (%)	Time (ms)	FLOPs (B)
Park et al. [111]	78.4	70.3	76.7	1.0	4.2
Yang et al. [112]	79.1	69.7	76.3	1.1	2.3
Xu et al. [113]	78.2	69.9	76.7	1.1	4.5
Liu et al. [114]	79.5	70.8	78.1	0.9	8.2
Zhu et al. [115] *	75.8	71.3	76.4	0.8	7.7
Zhao et al. [116] *	78.0	71.3	77.6	3.6	17.9
YOLOv5n	78.3	69.7	76.7	0.9	4.1
YOLOv5s	79.7	72.7	78.2	2.2	15.8
YOLOX tiny	76.8	69.8	75.7	2.7	15.4
YOLOv6n	77.8	69.8	77.6	1.3	11.3
YOLOv7 tiny	78.3	72.2	78.4	1.2	13.0
YOLOv8n	77.8	70.0	77.4	0.8	8.1
GELAN *	78.4	69.5	77.9	1.0	8.1
YOLOv9 *	78.8	70.3	78.0	3.5	18.1
YOLOv10n	78.3	69.5	77.1	0.9	8.2
ESFD-YOLOv8n (Proposed)	80.1	72.7	79.4	1.0	10.8

* To match YOLOv8n, the model’s size was reduced.

with the optimal values in bold, demonstrated the superior performance of ESFD-YOLOv8n. The proposed model achieved a precision of 80.1%, a recall of 72.7%, a [email protected] of 79.4%, and a remarkable inference time of 1.0 ms, highlighting its effectiveness and efficiency in object detection on the D-Fire dataset.

Table 6 presents the exceptional detection performance of the proposed ESFD-YOLOv8n model across all accuracy-related metrics. The proposed ESFD-YOLOv8n model outperformed the other methods, showing higher precision, recall, and [email protected] values (Figure 11).

In particular, ESFD-YOLOv8n achieves notable improvements in precision over the existing methods, exceeding [111,112,113,114,115,116], YOLOv5n, YOLOv5s, YOLOX tiny, YOLOv6n, YOLOv7 tiny, YOLOv8n, GELAN, YOLOv9n, and YOLOv10n by 1.7%, 1.0%, 1.9%, 0.6%, 4.3%, 2.1%, 1.8%, 0.4%, 3.3%, 2.3%, 1.8%, 2.3%, 1.7%, 1.3%, and 1.8%, respectively. Furthermore, it outperformed the [111,112,113,114], YOLOv5n, YOLOX tiny, YOLOv6n, YOLOv7 tiny, YOLOv8n, GELAN, YOLOv9, and YOLOv10n methods in recall by 2.4%, 3.0%, 2.8%, 1.9%, 1.4%, 1.4%, 3.0%, 2.9%, 2.9%, 0.5%, 2.7%, 3.2%, 2.4%, and 3.2%, respectively, achieving results comparable to YOLOv5s. Similarly, ESFD-YOLOv8n demonstrated remarkable improvements over the [111,112,113,114], YOLOv5n, YOLOv5s, YOLOX tiny, YOLOv6n, YOLOv7 tiny, YOLOv8n, GELAN, YOLOv9, and YOLOv10n methods in [email protected], surpassing other methods by 2.7%, 3.1%, 2.7%, 1.3%, 3.0%, 1.8%, 2.7%, 1.2%, 3.7%, 1.8%, 1.0%, 2.0%, 1.5%, 1.4%, and 2.3%, respectively.

Despite the slight increase in inference time and floating-point operations observed compared with the baseline model, resulting in a minor decrease in FPS, the significant gains in accuracy, with [email protected], precision, and recall increasing by 2.0%, 2.3%, and 2.7%, respectively, highlighting the value of the tradeoff.

These substantial advancements firmly establish ESFD-YOLOv8n as a state-of-the-art solution for smoke and fire detection in complex environments, maintaining real-time performance while significantly enhancing accuracy.

3.7. Qualitative Analysis between ESFD-YOLOv8n and the Original YOLOv8n Model

Figure 12 shows the prediction results from the ESFD-YOLOv8n and original YOLOv8n models.

Figure 12 confirms the superior detection performance of ESFD-YOLOv8n in real-world fire detection applications compared with the original YOLOv8n model. In particular, Figure 12b shows how the enhanced model prioritizes relevant information and effectively preserves features crucial for detecting small smoke plumes, a critical aspect of early fire safety. Furthermore, the original YOLOv8n model showed instances of missed detections (false negatives, Figure 12c,i,l) and incorrect identifications (false positives, Figure 12f), issues that are reduced significantly in the ESFD-YOLOv8n model.

4. Conclusions and Future Work

This paper introduced ESFD-YOLOv8n, a novel YOLOv8n-based model demonstrating significant improvements in smoke and fire detection accuracy and efficiency in early smoke and fire detection. Leveraging key technologies, such as ResBlock, WIoUv3 loss, C2fGELAN blocks, and transfer learning, ESFD-YOLOv8n outperformed the existing state-of-the-art methods on the D-Fire dataset, achieving an impressive mAP50 of 79.4% with a real-time inference time of 1.0 ms.

However, a thorough error analysis revealed several challenges. The dataset suffers from an imbalanced number of classes and a high number of background images (Figure 13). This imbalance leads to a relatively high difference in accuracy among classes, an increased false positive rate, and reduced generalization.

Additionally, model performance varies in different scenarios, including redundant boxes at far distances, lack of detection with similar backgrounds, and incorrect detection of similar objects, as shown in Figure 14.

To address the issue of redundant boxes at far distances, it is recommended to implement NMS and adjust anchor box scales. For the lack of detection with similar backgrounds, using a diverse training dataset, applying data augmentation techniques such as random cropping and color jittering, exploring alternative loss functions, and addressing class imbalance are suggested. To improve detection accuracy for similar objects, increasing object class granularity and incorporating contextual information are recommended.

Overall, recommendations for enhancing the model include fine-tuning hyperparameters, regularly evaluating the model on a validation set, and considering ensemble methods. In our future work, we will address these limitations to further refine the model. By overcoming these challenges, we aim to develop a more reliable and versatile smoke and fire detection model for practical applications by enhancing accuracy and reducing false alarms. This advancement is expected to improve safety measures and early warning systems, ultimately mitigating fire hazards and protecting communities. The promising results of ESFD-YOLOv8n lay the groundwork for further research in this critical field, paving the way for more sophisticated fire detection systems across various applications.