VBM-YOLO: an enhanced YOLO model with reduced information loss for vehicle body markers detection

Object detection model

Object detection models are primarily categorized: two-stage detectors and one-stage detectors. Classic two-stage detectors include R-CNN (Girshick et al., 2014), Fast R-CNN (Girshick, 2015), Faster R-CNN (Ren et al., 2015), while representative one-stage models include SSD (Liu et al., 2016), FCOS (Tian et al., 2020), and YOLO series (Redmon et al., 2016; Redmon & Farhadi, 2017, 2018; Bochkovskiy, Wang & Liao, 2020; Ultralytics, 2022; Li et al., 2023; Wang, Bochkovskiy & Liao, 2023).

Two-stage object detection models improve the accuracy and robustness of detection by first generating candidate boxes and then classifying and adjusting these candidates. R-CNN (Girshick et al., 2014), one of the earliest two-stage models, introduced convolutional neural networks (CNNs) to the object detection field. He et al. (2017) further enhanced this model by adding a branch for predicting object masks, leading to the development of Mask R-CNN. Cai & Vasconcelos (2018) proposed Cascade R-CNN, which achieves higher accuracy by cascading multiple detectors. Despite the methods above attaining high accuracy and superior performance, particularly for large objects and complex scenes, the processing of two independent network stages results in slower speeds.

In contrast, one-stage models directly predict object detection results from input data without undergoing multi-stage processing, thus achieving greater efficiency. OverFeat (Sermanet et al., 2013) was one of the earliest CNN-based one-stage models. Currently, RT-DETR (Zhao et al., 2023) is regarded as one of the most generalized and accurate one-stage object detection models in the DETR series. However, its extensive parameters and substantial training resource requirements hinder practical deployment. The YOLO series has been widely adopted in industrial applications due to its real-time detection capabilities. Among them, YOLOv8 (Terven, Córdova-Esparza & Romero-González, 2023) is recognized as one of the most advanced versions in the YOLO family. Through continuous iterations, recent versions including YOLOv9 (Wang, Yeh & Liao, 2024), YOLOv10 (Wang et al., 2024), YOLOv11 (Khanam & Hussain, 2024), and YOLOv12 (Tian, Ye & Doermann, 2025) have achieved significant improvements in both detection accuracy and efficiency while preserving real-time performance.

Although the YOLO series performs well on large-scale datasets, it may not achieve high precision in specific scenarios, often requiring increased computational resources to enhance accuracy. For the specific task of vehicle body markers detection, we propose the VBM-YOLO model to achieve a balance between accuracy and computational efficiency, maintaining high detection accuracy while using fewer computational resources than similar models.

Attention mechanisms

In human perception, attention mechanisms involve selectively focusing on relevant information, allowing individuals to disregard irrelevant stimuli. This process helps the brain capture crucial information and grasp essential details. In recent years, various research methods have effectively integrated this attention mechanism into the architecture of deep CNNs to enhance the performance of large-scale visual tasks.

Early work, such as SENet (Hu, Shen & Sun, 2018), focused on modeling channel relationships in feature maps by learning weights for each channel. Subsequently, the convolutional block attention module (CBAM) (Woo et al., 2018) was proposed. It enriches attention maps by extracting channel and spatial features through max-pooling and average-pooling. To establish connections within the dimensions, DANet (Fu et al., 2019) modeled global dependencies of channels and spatial dimensions through self-attention mechanisms. The convolutional triplet attention module (CTAM) (Misra et al., 2021) models cross-dimensional dependencies between spatial and channels, highlighting the importance of capturing cross-dimensional dependencies. The coordinate attention (CA) (Hou, Zhou & Feng, 2021) addressed the issue of positional information loss in 2D pooling by decomposing channel attention into two 1D encoding processes. Recently, the expectation-maximization attention (EMA) (Ouyang et al., 2023) further enhanced the CA module by improving the sequential processing method, avoiding the side effects of channel dimension reduction modeling, and achieving channel reshaping and cross-dimensional fusion.

However, most of the methods mentioned above have significant drawbacks. The majority of attention modules consider only variations within a single dimension. Although modules like CBAM and DANet account for both dimensions, they fail to leverage the interaction between the two dimensions to facilitate better fusion. Similarly, while CTAM considers cross-dimensional information, it does not establish a global dependency within the spatial and channel dimensions, leading to the loss of some critical information. In contrast, our CSCA attention module not only employs self-attention to model local and global semantic dependencies between spatial and channel dimensions but also achieves information interaction between the two dimensions.

Pyramid networks

In object detection, researchers have extensively studied the fusion of multi-level features. These approaches aim to enhance detection performance by integrating high-level and low-level information. The well-known feature pyramid network (FPN) (Lin et al., 2017) is a classic feature fusion method in deep learning-based computer vision. It utilizes a top-down pathway to merge features at each level, thereby improving the detection performance of objects of varying sizes.

Given the exceptional performance of FPN, several variants have been proposed for different scenarios, such as Path Aggregation Network (PANet) (Liu et al., 2018), Bi-directional Feature Pyramid Network (BiFPN) (Tan, Pang & Le, 2020), Asymptotic Feature Pyramid Network (AFPN) (Yang et al., 2023), Info-FPN (Chen et al., 2023), and Attentional Bidirectional Feature Pyramid Network (ABFPN) (Zeng et al., 2022). Each variant optimizes the fusion pathways of FPN in distinct ways. However, pyramid networks with up-and-down varying structures, like those mentioned above, require high-level or low-level features to pass through multiple intermediate-scale transformations. This process gradually degrades feature information. Moreover, mixing multiple target features within the network makes extracting information for each target layer challenging. To this end, we propose MSSFPN, a pyramid network with a clear hierarchy that directly integrates low-level and high-level information in a progressive manner, reducing the loss caused by intermediate transformations.

Deep supervision

Deep supervision has been widely applied to enhance the performance of deep CNNs. These techniques (Wang et al., 2015; Szegedy et al., 2015; Wu, Hong & Chanussot, 2022) typically involve adding auxiliary prediction layers to certain intermediate hidden layers of deep neural networks, thereby supervising the main network branches. This approach helps addresses issues such as gradient vanishing and slow convergence during the training of deep neural networks. Recent studies in segmentation networks (Qin et al., 2020; Qi, Wu & Chan, 2023) and object detection Wang, Yeh & Liao (2024) have also demonstrated that deep supervision mechanisms significantly improve network performance. However, in traditional deep supervision, if shallow layer supervision fails to retain critical information during training, it may lead to error accumulation can occur, hindering subsequent layers from retrieving necessary information. Furthermore, most deep supervision methods require participation throughout the training and inference phases, which may result in additional network overhead. To address these issues, we propose the AGB method, which enables each target layer to mitigate the problem of gradient information vanishing while avoiding extra inference overhead.

Overall architecture

To address the issues mentioned above, we propose a novel VBM-YOLO model for vehicle body markers detection. The inference architecture of VBM-YOLO is shown in Fig. 1. We introduced the CSCA mechanism, which employs self-attention to establish cross-spatial-channel attention within the backbone network. The MSSFPN replaces the original neck network. Furthermore, within the MSSFPN network, we utilized the improved C2f-Rep module to enhance the extraction of channel and spatial information. The designed AGB is incorporated during training to monitor and guide information flow during training, ensuring more effective supervision, as detailed in the “Auxiliary Gradient Branch” section.

The main modules utilized in the VBM-YOLO backbone network, including CBS, C2f, and SPPF. Moreover, like YOLOv8, VBM-YOLO employs decoupled heads. The CBS module comprises three submodules: convolution, batch normalization, and the SiLU activation function. The C2f module evenly splits the extracted features along the channel direction using a slicing operation. These split features undergo processing through concatenated Bottleneck modules, followed by cross-layer connections. The SPPF module is an efficient pooling module designed to extract and fuse high-level features. This module utilizes repeated max-pooling to extract as many high-level semantic features as possible during the fusion process. Each detection head has two branches: a regression branch and a classification branch, which perform regression and classification tasks, respectively. Each branch consists of two 3 × 3 CBS and a 1 × 1 2D convolution. The detection head module employs an anchor-free approach and a dynamic TaskAlignedAssigner for positive and negative sample allocation. The primary structures of these modules are illustrated in Fig. 2.

Cross-spatial-channel attention

Vehicles may not always be detected from the optimal angle. Especially in scenarios like those shown in Fig. 3, strong lighting and long distances can make vehicle body markers inconspicuous in the entire image. Additionally, convolution operations produce local receptive fields, which can cause features of targets with the same pixels to differ. Therefore, we establish global dependencies between features through the cross-spatial-channel attention (CSCA) mechanism, adaptively aggregating remote context information, highlighting inconspicuous object features, and addressing feature discrepancy issues.

As shown in Fig. 1, CSCA includes a spatial attention module, a channel attention module, a cross-dimensional module, and a gated fusion unit. CSCA first reduces the dimensionality of features and inputs them into the spatial attention module. The spatial attention module selectively aggregates features at each position through a weighted sum of features at all positions. This ensures that similar features are correlated regardless of their distance. Meanwhile, the channel attention module selectively emphasizes interdependent channel maps by integrating related features across all channel maps.

Specifically, as shown in Fig. 4, the spatial attention module first generates a spatial attention matrix that models the spatial relationships between any two pixels of the features. Next, matrix multiplication is performed between the attention matrix and the original features. Finally, element-wise summation is conducted between the resulting matrix and the original features to obtain the final representation reflecting remote context. The process of the channel attention module is similar to the spatial attention module, except that the first step is to calculate the channel attention matrix along the channel dimension. We use the gated fusion unit to reallocate the global feature information extracted by the spatial and channel attention modules, obtaining the critical global information that is most beneficial for each target layer. To achieve better feature representation across the spatial and channel dimensions, we fuse the interactive feature information generated by the cross-dimensional module with the gated fusion unit’s output through addition and convolution operations. We introduce a CSCA before each target layer to enhance its feature information.

Multi-scale selective feature pyramid network

To address the confusion and degradation of feature information caused by the previous pyramid structure, we propose the multi-scale selective FPN, as illustrated in the neck section of Fig. 1. MSSFPN employs a clear hierarchical structure for each target layer to avoid confusion. Specifically, it first fuses the low-level information from two target layers and then progressively integrates the higher-level information from deeper layers. This approach prevents information loss associated with top-down operations and mitigates conflicts from directly fusing low-level and high-level information with significant semantic gaps. Direct summation during feature fusion inevitably leads to information loss. Therefore, we introduce the ASFF (Liu, Huang & Wang, 2019) module to learn feature information from different target layers adaptively. Although ASFF handles the fusion of features at different scales, preserving the crucial feature information of the current layer is also essential. Thus, we propose the feature selective fusion (FSF) module, which selectively integrates original and multi-scale information better to capture the crucial information of the current target layer. In the FSF module, we first perform global average pooling on both the original and fused features to obtain two sets of pooled feature information. These sets are then concatenated along the channel dimension and processed using softmax and sigmoid, respectively, to generate two sets of weights representing the importance levels for the current target layer. These weights are multiplied with the original features, and the resulting features are fused to produce information beneficial for the current layer.

Additionally, our research finds that the BottleNeck module in the original C2f structure of YOLOv8, which uses two consecutive 3 × 3 convolutions, leads to the loss of some channel fusion information. Consequently, within the MSSFPN, we replace the BottleNeck module with the RepViT Block module from RepViT (Wang et al., 2023) in C2f. The RepViT Block module, with its multi-branch structure and 1 $\times$ 1 and 3 $\times$ 3 depthwise convolutions, retains channel information and enhances spatial information while having fewer parameters than the original BottleNeck module. Furthermore, we utilize the reparameterization technique from the RepViT Block to reduce GFLOPs by 1.3 during inference.

Auxiliary gradient branch

In object detection algorithms, information from each target layer is transmitted from the preceding layer to the current one, often resulting in information loss. This issue becomes more pronounced as the network depth increases. Therefore, it is crucial to maintain the integrity of each target layer’s information, which requires the complete transmission of information from the previous layer to the current one, both in forward and backward propagation. Given our improvements to the network structure, we aim to preserve the gradient information of each target layer from the perspective of backward propagation. As illustrated in Fig. 5, to ensure consistency in information during forward and backward propagation, the backbone structure in the AGB is kept consistent with the backbone in VBM-YOLO. Each target layer in VBM-YOLO is connected to all corresponding preceding target layers in the AGB, enabling it to receive gradient information from the previous layer and thus maintain its information integrity. For instance, during training, to ensure the completeness of gradient information in the third target layer of VBM-YOLO, the third target layer is sequentially connected to the first to third target layers of the AGB branch. During backpropagation, the gradient information from the first to third layers in the AGB is sequentially propagated to the third target layer of VBM-YOLO, thus ensuring the integrity of the third target layer’s information.

When employing the AGB, an additional branch is added behind each CBS module in VBM-YOLO to connect a Link module. The Link module, composed of 1 × 1 convolutions, sequentially converts the input feature maps to match the channel number of the target layers in the AGB. These converted feature maps are then sequentially assigned to the Fuse module in the AGB for merging. The Fuse module upsamples the features from different target layers and sums them to obtain fusion information on different layers. Moreover, the AGB is used only during training and removed during inference to reduce computational overhead.

Datasets

This study employs portable mobile devices to collect vehicle images requiring body marking detection. From an extensive image pool, we curated a representative dataset containing 7,164 images-comprising 3,582 right-rear 45° vehicle views and 3,582 left-front 45° vehicle views. As shown in Fig. 3, to better simulate real-world detection environments and enhance model robustness, the dataset intentionally includes 122 images with significant overexposure and 132 images captured from substantially long distances. As shown in Fig. 6, this dataset contains six detection targets: rear reflective signs, side reflective signs, rear guards, side guards, rear signs, and triangle signs. Specifically, the rear right side images contain rear reflective signs, side reflective signs, rear guards, side guards, rear signs, and triangle signs. The front left side images contain side reflective signs and side guards.

Ground truth annotations were generated using the LabelImg software. The dataset was divided into a training set of 6,164 images and a validation set of 1,000 images, with an equal 1:1 distribution of right rear and left front images in both sets. Figure 7 illustrates the distribution of labels in the vehicle body markers dataset used in this study. To further verify the generalization of our model across different tasks, we conducted tests on two public datasets: the PASCAL VOC (Everingham et al., 2010) dataset and the D-Fire (de Venâncio, Lisboa & Barbosa, 2022) dataset, as shown in Fig. 7 for label distribution. The PASCAL VOC dataset, including the 2007 and 2012 versions, contains 20 categories of detection objects. In our experiments, we used 16,551 images from the training and validation sets of VOC2007 and VOC2012 for training and 4,952 images from the VOC2007 test set for validation. The D-Fire dataset, which consists of images of fire and smoke events, has two categories: smoke and fire, with 21,527 images. We used the official split of 17,221 images for training and 4,306 images for validation.

Experimental setup

The experiments were conducted on an NVIDIA GeForce RTX 4090 (24 GB) GPU, utilizing the environmental dependencies of PyTorch 2.0.1, Python 3.9.17, and CUDA 11.7. During training, we used the SGD optimizer with a batch size of 16. The hyperparameters for SGD were set as follows: momentum at 0.937, initial learning rate at 0.01, and weight decay at 0.0005. All input images were uniformly scaled to $640\times 640$ pixels. As depicted in Fig. 8, the training process of the YOLOv8s model for our collected vehicle body markers dataset exhibited stability in all evaluation metrics after 125 to 200 epochs. Therefore, the maximum epoch count was set to 200 for models trained using our collected dataset. For the PASCAL VOC dataset and the D-Fire dataset, the maximum epoch count was set to 300 to ensure adequate training and optimal performance on these datasets. In this study, all models were trained from scratch without using pre-trained weights.

Data augmentation

To enhance the model’s generalization capability, we systematically applied a range of data augmentation techniques. We introduced the Mosaic augmentation method, which generated composite training samples by randomly selecting, stitching, and cropping four images. This approach not only effectively improved the detector’s adaptability to various backgrounds through multi-background fusion but also allowed the Batch Normalization (BN) layers to learn the feature distributions of four images simultaneously within a single batch, thereby significantly enhancing the model’s feature extraction performance.

Regarding the fundamental augmentation strategies, we employed a variety of multi-dimensional geometric transformations: (1) mirror flipping, which produced mirrored samples via horizontal and vertical flips to increase the diversity of the training data; (2) orthogonal rotation, which applied rotations in multiples of 90° to enhance the model’s resilience to changes in object orientation; (3) random cropping, which involved sampling local regions to force the model to focus on the most salient features of the targets; and (4) angle rotation, which performed rotations at arbitrary angles in combination with a dynamic bounding box coordinates correction algorithm, ensuring precise spatial alignment of the detection boxes with the rotated images. The integration of these methods significantly improved the model’s robustness against variations in target scale, orientation, and background.

Evaluation metrics

In this study, the model size is evaluated based on its number of parameters. Precision (P), recall (R), and mean average precision ( $mAP$) were employed to evaluate the detection performance of the proposed model. P indicates the probability that a detected object is correctly identified among all detected objects. R denotes the probability of correctly identifying positive samples among all positive samples. Average precision (AP) measures the accuracy for a specific category and is determined by the area under the precision-recall (P−R) curve. $mAP$ is the mean of the (AP) values for all categories, indicating the overall accuracy of the model. A higher $mAP$ indicates more accurate detection outcomes. The following equations provide detailed explanations of these metrics:

(12) $$P=\frac{TP}{TP+FP}$$

(13) $$R=\frac{TP}{TP+FN}$$

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

(15) $$mAP=\frac{{\sum }_{n=1}^{N}AP(n)}{N}$$where TP denotes the number of true positive samples, FP denotes the number of false positive samples, and FN denotes the number of false negative samples. N represents the number of classes in the dataset.

Furthermore, $mAP@0.5$ refers to the mean average precision when the Intersection over Union (IoU) threshold is set to 0.5, while $mAP@0.5:0.95$ indicates the mean average precision across IoU thresholds ranging from 0.5 to 0.95 with increments of 0.05.

Comparison experiments

In this section, we conduct experiments on three datasets and compare VBM-YOLO with mainstream one-stage and two-stage object detection models to evaluate its effectiveness.

Table 1 presents the performance comparison between current mainstream models and our proposed VBM-YOLO model on the vehicle body markers dataset. The proposed VBM-YOLO model exhibits superior performance compared to other models. Compared with two-stage models such as Faster R-CNN and SSD, VBM-YOLO shows significant advantages in model size and mAP. Although our VBM-YOLO model is larger than the original YOLOv8s, it achieves a notable improvement in mAP, with increases of 2.3% in [email protected] and 4.3% in [email protected]:0.9. Moreover, in comparison with the larger YOLOv8m model, VBM-YOLO not only maintains a smaller size but also attains higher [email protected] and [email protected]:0.9, with gains of 1.3% and 2.2%, respectively. The table shows that, while YOLOv7-tiny has the smallest number of parameters, its accuracy is substantially lower compared to the VBM-YOLO model. We also compared our model with some of the latest object detection models of similar sizes, including RT-DETR based on the ResNet18 backbone network, YOLOv9-C, and GELAN-C. It should be noted that GELAN-C is an efficient layer aggregation network proposed in YOLOv9. However, compared to YOLOv9-C, which has the highest accuracy and the largest size among the three, our VBM-YOLO model reduces the number of parameters by 28% while still improving [email protected] and [email protected]:0.9 by 1.4% and 2.4%, respectively. Comparative evaluations with recently released YOLOv10, YOLOv11, and YOLOv12 demonstrate VBM-YOLO’s [email protected] improvements of 3.7%, 2.1%, and 1.6%, and [email protected]:0.9 gains of 5.1%, 5.0%, and 3.6% respectively, while maintaining substantially lower computational costs and higher FPS. These results in the table indicate that our model achieves the best balance between accuracy and resource consumption compared to other models. The PR curve in Fig. 9 further illustrates the improved detection performance of VBM-YOLO compared to YOLOv8s across various datasets. Despite the challenges posed by intense lighting and varying distances, which reduce detection accuracy for side reflective signs, our model still outperforms YOLOv8s by 1.4% in detecting these markers. The results from the other two datasets also clearly demonstrate that VBM-YOLO outperforms YOLOv8s in most detection categories. Under high-intensity illumination conditions, distant pixels are frequently surrounded by saturated light regions, resulting in diminished contrast differentiation between adjacent pixels. This photometric interference may cause detection models to overlook subtle variations between pixels during feature extraction or even completely ignore the feature information from these light-affected regions. To address this limitation, our detection framework incorporated CSCA module that explicitly preserves illumination-aware feature distinctions between proximal and distant pixels throughout the feature hierarchy. As illustrated in Fig. 10, this mechanism achieved significant enhancement in the model’s detection performance.

To further validate VBM-YOLO’s generalization performance, we conducted comparison experiments on the public PASCAL VOC and D-Fire datasets. As shown in Tables 2 and 3, comparison experiments were conducted on the publicly available PASCAL VOC and D-fire datasets to evaluate the generalization capabilities of our proposed model against other mainstream object detection models.

For results in the PASCAL VOC dataset, though our VBM-YOLO model is specifically designed for detecting vehicle body markers, it also demonstrated superior performance, surpassing the baseline YOLOv8s model and other earlier YOLO versions. VBM-YOLO achieved higher accuracy than RT-DETR, which is based on the ResNet18 backbone, despite having fewer parameters. While YOLOv9-C showed higher accuracy than VBM-YOLO, it required 7.1 million additional parameters.

For results in the D-fire dataset, while YOLOv10-M achieved higher accuracy than VBM-YOLO on the PASCAL VOC dataset, its accuracy on the D-fire dataset was significantly lower than that of VBM-YOLO. Furthermore, while YOLOv9-C had a higher [email protected], VBM-YOLO achieved a higher mAP@50:95 with fewer parameters. Excluding YOLOv9-C, which has substantially more parameters than VBM-YOLO, our model achieved the highest accuracy on both metrics in the D-fire dataset compared to other models.

The above analysis indicates that the proposed VBM-YOLO model not only excels in vehicle body markers detection but also exhibits strong generalization capabilities across other detection tasks. To visually demonstrate the performance advantages of our VBM-YOLO algorithm, Fig. 11 compares the detection capabilities of VBM-YOLO, YOLOv8, YOLOv9, and YOLOv10. Although there are significant differences in the number of model parameters between YOLOv8 and YOLOv9, their final detection performances do not exhibit notable distinctions. The recently released YOLOv10 has made substantial progress in mitigating latency issues. However, its specific optimizations for the COCO dataset and lightweight design may somewhat limit its efficacy in vehicle body markers detection. Notably, all three models exhibit a certain degree of missed detections. VBM-YOLO, benefiting from techniques designed to address information loss, demonstrates a significant advantage in reducing missed detections compared to the other algorithms.

Ablation experiments

The ablation experiments were conducted to verify the effectiveness of each improvement module in the VBM-YOLO model on the vehicle body markers dataset. The results are shown in Table 4. Initially, the CSCA module was added individually before the SPPF module in the baseline model. It was observed that [email protected] and [email protected]:0.95 increased by 1% and 0.4%, respectively, compared to the baseline model. Subsequently, a separate ablation experiment was conducted on the MSSFPN, replacing the original baseline model’s neck network. This led to significant improvements, with [email protected] and [email protected]:0.95 increasing by 2% and 3.3%, respectively. These results demonstrate the effectiveness of the proposed MSSFPN feature pyramid architecture for feature fusion and extraction.

Additionally, we found that employing the ASFF module with a sigmoid function in MSSFPN produced better results than using the ASFF module with a softmax function. Both [email protected] and [email protected]:0.95 increased by 0.1%. An ablation experiment was also conducted on the C2f module within the MSSFPN. It was found that using the C2f-Rep model, which incorporates channel fusion feature information, resulted in a 0.1% decrease in [email protected] but a 0.3% increase in [email protected]:0.95. Therefore, the MSSFPN was implemented with the ASSF module using the sigmoid activation function and the C2f-Rep module.

To explore the effect of using the CSCA module and MSSFPN simultaneously, the CSCA attention mechanism was added before each target layer in the VBM-YOLO model’s backbone, and the MSSFPN was used in the neck network. A significant improvement in accuracy was observed compared to using the CSCA and MSSFPN individually. Compared to the initial baseline model, [email protected] and [email protected]:0.95 increased by 2.4% and 3.9%, respectively. Furthermore, to alleviate the loss of gradient information for each target layer during backpropagation, the AGB branch was used to supervise training and removed during inference. It was observed that although [email protected] decreased by 0.1% compared to not using AGB supervision, [email protected]:0.95 increased by 0.4%. Therefore, we decided to use the AGB branch to assist in training our model.

In addition, to further validate the superiority of the proposed CSCA attention mechanism over other widely used attention mechanisms in vehicle body markers detection, we replaced CSCA in the VBM-YOLO with attention mechanisms employing similar approaches, such as iRMB (Zhang et al., 2023), Biformer (Zhu et al., 2023), EMA (Ouyang et al., 2023), and DilateFormer (Jiao et al., 2023). It is important to note that the AGB branch was not used during these experiments. The results, as shown in Table 5, clearly demonstrate that our attention mechanism achieves the best balance between accuracy and computational resources when incorporated into the model, compared to other attention modules.

We used heatmaps, as illustrated in Fig. 12, to visually observe the performance of each attention mechanism. Compared to our CSCA attention mechanism, the heatmaps for other attention mechanisms exhibit missing or lighter thermal information in the regions requiring detection. In contrast, the CSCA heatmap shows significantly darker color distribution in the areas that need to be detected. These results indicate that the CSCA attention module captures key information in the features more effectively than other attention modules.