2.1 Overall Architecture

Figure 1 Overall architecture of the proposed object detector.

To mitigate overfitting, we monitor both training and validation performance. We applied Dropout and L2 regularization during training, and utilized data augmentation techniques to improve generalization. This helps ensure that the model does not overfit to the training data.

The overall architecture of our proposed object detection method is illustrated in Figure 1, where the purple dotted box highlights some basic modules within the architecture, providing a detailed representation of their specific functions. The ELAN module and BRA module within the overall structure will be explained in detail in Sections 2.2 and 2.3. Initially, an input remote sensing image is resized to a size of 640x640. It is then fed into the backbone network, where image features are extracted using the Elan module. The Head network generates three layers of feature maps with different sizes, corresponding to large, medium, and small targets. Subsequently, the prediction results are obtained through the Rep and Conv processes, which involve the fusion of multi-scale features to enhance detection accuracy. Finally, we modify the network’s weights by comparing the predicted bounding boxes with the ground truth boxes using a loss function. This iterative process is repeated until the final prediction result is obtained. The CBS modules of three colors correspond to the convolution basic units of the different convolution kernels and strides and consist of convolution, batch normalization, and SiLU activation functions. SPPCSPC is a spatial pooling pyramid composed of convolution and max pooling. Its main function is to avoid image distortion caused by cropping and zooming operations on the image area. At the same time, it solves the problem of repeated feature extraction of convolutional neural networks, which greatly improves the speed of generating region proposals and saves computational costs.

2.2 Adaptive Input Backbone Network

Since the picture will lose a lot of information during the feature extraction process, YOLOv7 integrated the Cross Stage Partial Darknet-53 (CSPDarknet-53) structure into the backbone network ELAN module of YOLOv7, as shown in Figure 2, to ensure that the network can extract features better. The CSPDarknet-53 architecture is an improvement to the classic Darknet-53 backbone network, introducing Cross-Stage Partial connections to allow interaction between bottom-level and top-level features to improve feature propagation and information mobility. The design of this architecture makes YOLOv7 perform well in multi-scale target detection tasks, and is suitable for object detection of various sizes and shapes. Furthermore, by performing detections on feature maps at different levels, the network can better adapt to various object sizes, improving the model’s performance while maintaining a relatively low parameter amount. We propose that the extended ELAN (ELAN-H) approach preserves the original architecture’s gradient transmission path entirely. Instead, it employs group convolution to augment the cardinality of the newly added features and amalgamates them from various groups using a shuffle and merge cardinality technique.

In order to better extract features and allow the network to learn deformed objects, we replaced all the convolutions with the original convolution kernel size 3 in the last two ELAN modules of the backbone network with DCNv2, which can be more adaptable Object deformation, and also solves the problem that when DCNv1 expands the area of interest, irrelevant areas affect the performance of the network. This targeted use of DCNv2 represents one of the core innovations of YOLOv7-bw, specifically designed to improve dense and small target detection performance. In general, today’s ELAN backbone network is more adaptable to targets of different sizes and deformations, and has more powerful feature extraction capabilities. Compared with ELAN module, in ELAN-H module of the detection head network, we chose to maintain the original YOLOv7 design and did not replace the convolution with DCNv2, the structure of ELAN and ELAN-H is shown in Figure 2.

Figure 2 The structure of ELAN and ELAN-H module.

2.3 Bi-level Routing Attention (BRA) module

The primary objective of the self-attention mechanism is to enhance the network’s concentration on pivotal areas. Several self-attention modules mentioned earlier incorporate predetermined sparse patterns, which are manually crafted. Nevertheless, although different strategies are employed to merge or select key and value tokens, these tokens remain query-independent, meaning they are shared among all queries. However, queries associated with distinct semantic domains prefer different key-value pairs. Therefore, enforcing the same set of tokens for all queries may not be optimal. In contrast to other attention modules, BRA (Query-Aware Sparse Attention) represents a dynamic attention mechanism. Its objective is to empower each query to concentrate on a specific subset of the most semantically relevant key-value pairs. The core concept of BRA involves an initial filtration process at the regional level to eliminate the least relevant key-value pairs, thus preserving only a limited number of routing regions. Subsequently, fine-grained label-to-label attention is employed on the combination of these routing regions. By solely involving dense matrix multiplication, BRA performs satisfactorily while ensuring high computational efficiency. The specific steps can be roughly divided into the following three parts:

(1) Region division and input projection

Input a two-dimensional feature map, $𝑿\in {ℝ}^{H\times W\times C}$. First, it is divided into $S\times S$ non-overlapping regions, where each region contains $\frac{HW}{S^2}$ feature vectors, reshape X to ${𝑿}^r\in {ℝ}^{S^2\times \frac{HW}{S^2}\times C}$. Then we derive query Q, key K, value V, so the linear projection is:

where ${𝑾}^q,{𝑾}^k,{𝑾}^v\in {ℝ}^{C\times C}$, are the projection weights of query, key, and value, respectively.

(2) Area-to-area directed graph routing

Based on the first step, we find the engagement relationship by constructing a directed graph. First by applying the mean of each region to Q and K separately, get ${𝑸}^r,{𝑲}^r\in {ℝ}^{S^2\times C}$, then calculating the adjacency matrix ${𝑨}^r$ of the interregional correlation of ${𝑸}^r$ and ${𝑲}^r$ :

The adjacency matrix ${𝑨}^r\in {ℝ}^{S^2\times S^2}$ measures the degree of semantic correlation between two regions. Next, only the top k connections of each region are kept to prune the correlation graph. Specifically, the routing index matrix ${𝑰}^r\in {\mathbb{N}}^{S^2\times k}$ is used to save the indexes of the top k connections row by row:

where the $i$-th row of matrix ${𝑰}^r$ comprises the indices representing the top k most pertinent regions related to the i-th region.

(3) Token-to-token attention mechanism

We can implement fine-grained label attention using the region-to-region routing index matrix denoted as. Each query token within region i examines all key-value pairs situated in the collective set of k routing regions. The initial step entails the collection of the key and value tensors:

where ${𝑲}^g$ and ${𝑽}^g$ are the tensors of the aggregated key and value, and then use the attention operation on the aggregated key-value pairs:

A context enhancement term $LCE\left(𝑽\right)$ is introduced here, the function $LCE(\bullet )$ is parameterized with depth-wise separable convolution. Furthermore, we configured the convolution kernel size to be 5.

The BRA algorithm generally leverages sparsity to bypass the computation of the least significant areas by gathering key-value pairs within the top k pertinent windows. This approach exclusively utilizes GPU-compatible dense matrix multiplication, as depicted in Figure 3, where mm represents matrix multiplication.

Figure 3 Leveraging sparsity for improved computational efficiency in the BRA Algorithm.

The BRA module is incorporated into the YOLOv7 network. Initially, it directs the feature map obtained from region extraction to its respective regions and employs token-to-token attention to obtain a novel output feature map. Given that this operation involves the feature graph, and since the Spatial Pooling Pyramid (SPP) is designed to mitigate image distortion resulting from image processing and repeated feature extraction, it is determined to insert the BRA module after the SPP stage. The BRA module, as one of YOLOv7-bw’s primary innovations, significantly enhances the network’s capability to detect dense small targets by adaptively emphasizing relevant spatial features.

2.4 Optimized BBR Loss

Figure 4 Example of the BBR loss [53].

The loss function used in YOLOv7 for bounding box regression is CIoU [54], which increases the consideration of aspect ratio consistency based on the normalized length of the center point connection. However, it solves the problem that the anchor box cannot be optimized when the negative gradient $\frac{\partial R\mathrm{DIoU}}{\partial W_g}$ and $\frac{\partial L\mathrm{IoU}}{\partial W_g}$ offset. However, it is inevitable that numerous inferior anchor boxes will be produced while making predictions, thus incorporating geometric indicators like aspect ratio or distance will exacerbate the punishment for subpar anchor boxes, consequently diminishing the model’s capacity to generalize. An ideal loss function ought to diminish the impact of geometric metrics when the anchor box aligns closely with the target box, without excessively intervening in the training process. Here, the distance attention is constructed with the distance metric, breaking through previous BBR losses which that relied on additive losses, and the two are multiplied to obtain WIoUv1 with a two-layer attention mechanism:

where $R\mathrm{𝑊𝐼𝑜𝑈}\in [1,e)$ will significantly scale up the LIoU of normal quality anchor boxes, $L\mathrm{𝐼𝑜𝑈}\in [0,1]$ will significantly reduce the RWIoU of high-quality anchor boxes. In order to prevent RWIoU from producing gradients that hinder convergence, Wg, Hg are separated from the computational graph in Figure 4 (superscript * indicates this operation).

Although WIoUv1 can be applied to various scenarios, it doesn’t effectively address the issue of blurred targets in remote sensing images. To enhance our focus on blurred targets, we adopt WIoUv3, which includes a focusing mechanism by incorporating a gradient gain calculation method (focus coefficient) based on v1. We replace the CIoU loss function in YOLOv7 with WIoUv3 to better handle the detection of blurred targets in remote sensing images. WIoUv3 defines the outlier to describe the quality of the anchor box, which is specifically defined as:

where $\overline{L_{\mathrm{𝐼𝑜𝑈}}}$ represents the exponential running average with a momentum parameter m. Small outlier indicate high-quality anchor boxes and are assigned a low gradient gain. Similarly, anchor boxes with significant outliers also receive a low gradient gain, which effectively prevents low-quality examples from generating large, detrimental gradients. This approach ultimately directs the bounding box regression to prioritize anchor boxes of normal quality. A non-monotonic focusing factor is constructed using $\beta$ and applied to WIoUv1:

where $r$ denotes the gradient gain, $\alpha$ and $\delta$ refer to the artificially set hyperparameters, because r is dynamic, the criteria for dividing the anchor box mass are also dynamic in WIoUv3. This dynamic approach allows WIoUv3 to adapt its gradient gain allocation strategy to best suit the current situation at any given moment.

Figure 5 The relationship between $\beta$ and the gradient gain r is governed by the hyperparameters $\alpha$ and $\delta$.

After we plot the influence of several different groups of $\alpha$ and $\delta$ on the outlier $\beta$ and gradient gain r, as shown in Figure 5, the larger $\alpha$ is, the further to the left the peak of the curve is and the smoother the curve is; the larger $\delta$ is, the larger the peak of the wave is. The highest gradient gain for detectors with good performance should be more appropriate when the outlier degree is between 1 and 2. On the contrary, when applied to detectors with poor performance, more attention should be paid to the better anchor boxes of the fit, that is, the highest gradient gain should be when outliers are low. As for a detector like YOLOv7, which has a good detection rate, we hope to have the highest gradient gain r when the outlier $\beta$ is between 1 and 2, and, at the same time, have a lower gradient gain when the outlier is high. It can be seen that the blue curve has better performance in the low and high outlier. There is a small gradient gain at all times, so the loss function pays more attention to the anchor boxes of ordinary quality, and we finally select hyperparameters $\alpha$=1.9 and $\delta$=3 to apply to the final experiment (the experiment of hyperparameter determination is given in Section 3.2). Simultaneously, to avoid retaining low-quality anchor boxes during early training, we set the initial value $\overline{L_{\mathrm{𝐼𝑜𝑈}}}=1$, so that when LIoU=1, it has the largest gradient gain. The adoption of WIoUv3 as a bounding box regression loss function represents another key innovation of YOLOv7-bw, effectively addressing localization inaccuracies for densely clustered and blurred targets in remote sensing imagery.

2.5 Preprocessing operations and parameters settings

At the input of YOLOv7-bw, we used mosaic data enhancement, anchor frame adaptive calculation and image adaptive scaling to preprocess the input data. Remote sensing images mostly contain small targets, and usually the detection accuracy is much lower than that of large targets. Therefore, the mosaic data enhancement method was selected and 4 pictures were spliced by random scaling, cropping, and arrangement. This approach can greatly enrich the data set, especially random scaling, which adds many small targets, making the network more robust. At the same time, the data of 4 pictures is directly calculated, so the Mini-batch size does not need to be very large. A single GPU can achieve better results, thus reducing the need for GPU calculations.

What’s more, in order to predict anchor boxes more accurate, the clustering method is used to adaptively calculate the data in the training set to obtain a set of optimal anchor box values. This set of optimal anchor boxes is used for fine-tuning in subsequent training. It can better adapt to the size of the target in the data set and get better prediction results.

In order to make our YOLOv7-bw perform better on remote sensing image data sets, we made a series of settings for the network parameters. These include training for 300 epochs; the batchsize is set as 16; resizing the input image size to 640*640; initial learning rate 0.01; final learning rate 0.1, etc.

3.1 Experimental evaluation metrics

In order to validate the integrity of the model, we utilized various evaluation metrics, including TP (True Positive), TN (True Negative), FP (False Positive), FN (False Negative), Precision, Recall, AP (Average Precision), precision-recall (P-R) curve, and mAP (mean Average Precision). These indicators provide insights into the relationship between the sample’s category and the model’s predicted results:

The precision is the percentage of the predicted true positive samples in the entire predicted positive samples, and the calculation result is:

where $recall$ is the percentage of the predicted true positive samples in the entirely true positive sample, and the calculation result is:

where P-R curve illustrates the correlation between precision and recall. By plotting recall on the x-axis and precision on the y-axis, a specific curve corresponding to a particular category can be generated. Furthermore, the enclosed area of this curve represents the average precision (AP) value for that category.

Moreover, the mAP indicator serves as a comprehensive evaluation metric for target detection algorithms applied to a specific dataset. It entails calculating the average of AP values across all categories. Currently, mAP stands as the paramount indicator for assessing model performance, and it can be mathematically expressed as:

3.2 Comparative experimental results

A series of comparative experiments were conducted to verify the effect of the improved YOLOv7 in this paper. The DIOR remote sensing image dataset was used, meticulously gathered by proficient individuals in the realm of Earth observation interpretation. This dataset encompasses 23,463 remote sensing images, comprising 190,288 manually annotated object instances. The dimensions of the images in the Dataset are 800×800 pixels, with a spatial resolution ranging from 0.5m to 30m. These instances are delineated by meticulously positioned bounding boxes across 20 prevalent object categories. The categories in datasets encompass airplanes, airports, baseball fields, basketball courts, bridges, chimneys, dams, highway service areas, highway tollbooths, ports, golf courses, ground track fields, overpasses, ships, stadiums, storage tanks, tennis courts, train stations, vehicles, windmills. The corresponding count for each category is presented in the Figure 6. It can be seen that there are a large number of vehicles and ships in the dataset but a small number of train stations and express-toll stations. The detection effect trained by such a dataset consistent with the actual number of objects in real life is more helpful for real-world tasks.

Figure 6 Targets and quantities description in the database.

For the experiments stage, we have constructed a cutting-edge deep learning model using the widely adopted deep learning compiling framework of Pytorch. All our experiments were conducted on a high-performance workstation equipped with an Intel Xeon E5-2643 v3 CPU and eight Nvidia Tesla P40 graphics cards, each with 24GB of memory. For the training phase, we utilized the YOLOv7 model as a foundation. Throughout the process, we conducted 300 training rounds, with a batch size of 16.

The commonly used algorithms in the field of target detection are selected for comparison with the improvements in this paper, including classic algorithms such as Faster R-CNN[22], SSD[29], RetinaNet[55], and CornerNet[56]. At the same time, a series of YOLO methods are also selected to verify the comparison model. All comparative methods were carried out under similar pre-training and testing conditions. The object identification results of different types and the whole in the database are shown in Table 1. Among them, the first row of the table represents the detection algorithm used, the first column represents different categories, and the numerical result is the AP for this category.

The last row is the average AP for all classes in the Dataset. Our YOLOv7-bw greatly improves vehicle and airplane compared with other algorithms. The reason is that vehicles and airplanes belong to the small target in remote sensing images, and vehicles are more likely to cluster and appear densely, which proves that YOLOv7-bw performs better on small and dense targets. For the target like the category of golffield, which belongs to the large target in the remote sensing image, although the improvement is not much, it has not decreased, which proves that our algorithm takes into account the large target at the same time, and has a certain degree of robustness. In addition to comparing the mAP of these classic algorithms, we also compared the precision and recall of our YOLOv7-bw with these algorithms, which obtain the highest precision and recall simultaneously, as shown in Figure 7.

Figure 7 Precision, Recall and mAP0.5 comparison histogram.

Finally, in evaluating the Dataset, we have carefully chosen a few images to compare the actual outcomes of the SSD, the YOLOv7 source code and our improved YOLOv7-bw version. The results of this comparison can be observed in Figure 8. The first image represents the original picture from the Dataset, followed by the SSD effect demonstration in the second image and the heat map generated by SSD in the third image, YOLOv7 effect demonstration and the heat map in the fourth and fifth image, our YOLOv7-bw effect demonstration and heat map in the sixth and final image.

Image (a) in Figure 8 shows that YOLOv7-bw produces fewer errors compared to YOLOv7 and More ships detected than SSD. In image (b), although YOLOv7-bw fails to detect several vehicles, which still surpass the SSD and 15 vehicles detected by YOLOv7. Moreover, the heat map indicates that YOLOv7-bw provides better accuracy and focus regarding vehicle positioning (the heat map generated by SSD and YOLOv7 can’t even locate the vehicle target). According to image (c), SSD only detected some airplane, although YOLOv7 detected more airplanes. Still, there were some false detections, only YOLOv7-bw, more airplanes were detected and no false detections occurred. In image (d), The detection rate of YOLOv7-bw is much higher than that of SSD and YOLOv7, and the focus area of the heat map is more concentrated on the target to be detected. To quantitatively assess the heatmap attention distribution, we computed the Mean Attention Score (MAS) within ground-truth bounding boxes. The MAS for YOLOv7-bw reached 0.78, compared to 0.63 for YOLOv7 and 0.51 for SSD, demonstrating that our proposed method effectively enhances target-focused attention.

Figure 8 Detection results of SSD, YOLOv7 and YOLOv7-bw. 1) the original picture; 2) the results by SSD; 3) a heat map generated by SSD; 4) the results by YOLOv7; 5) a heat map generated by YOLOv7; 6) the results by the proposed YOLOv7-bw; 7) a heat map generated by the proposed YOLOv7-bw.(a)ship; (b)vehicle; (c)airplane; (d)tenniscourt and vehicle.

Overall, as evident in Table 1, it’s clear that the YOLOv7-bw algorithm we introduced achieves the highest mAP at an IoU threshold of 0.5. This indicates that the our YOLOv7-bw algorithm enhances the detection performance of remote sensing images, resulting in an overall improvement. The model becomes competitive, especially in optimizing the detection of small and densely packed targets, effectively meeting the requirements for real-world remote sensing image detection in food supply applications.

Additionally, we evaluated the inference speed of the proposed YOLOv7-bw model to verify its practical applicability. On the same hardware environment (Nvidia Tesla P40 GPU), YOLOv7-bw achieves an average inference time of 8.2 ms per image, corresponding to approximately 122 FPS, which demonstrates its capability for real-time deployment. Despite the introduction of BRA and DCNv2 modules, the model maintains competitive computational efficiency compared to the baseline YOLOv7.

The experimental results show that YOLOv7-bw performs well in terms of both detection accuracy and speed. To evaluate for overfitting, we compared the performance of the model on the training and validation sets. The results showed that the performance gap between the training and validation sets was minimal, indicating that overfitting was not a concern. Regularization techniques such as L2 regularization and Dropout, as well as data augmentation, contributed to this result by ensuring the model generalizes well to new data.

3.3 WIoU hyperparameters experiment

To assess the best hyperparameters for implementing WIoUv3, we conducted a comprehensive set of experiments using the PyTorch framework. From the extensive categories available in the MS-COCO Dataset, we carefully chose 20 categories and used 28,000 images for training and 1,200 images for validation. For the model architecture, we selected the YOLOv7. During the training phase, our models underwent rigorous training for 120 epochs, utilizing a batch size of 32, and we applied different versions of the BBR loss function. The experimental results are eloquently presented in Table 2. Among them, mAP(50), mAP(75), and mAP(95) respectively represent the values of the map obtained when the IOU value is greater than 0.5, 0.75, and 0.95 in all categories.

According to the data presented in Table 2, it is evident that the effectiveness of WIoUv3, incorporating dynamic non-monotonic focusing, surpasses that of WIoUv1, which solely employs a focusing mechanism. At the same time, when $\alpha$=1.9 and $\delta$=3, WIoUv3 has the best performance under different IoU thresholds, which are 66.57, 57.03, and 49.81, respectively. Ultimately, we decided to use $\alpha$=1.9, $\delta$=3 as the final hyperparameter setting to participate in the experiment after.

3.4 Ablation experimental results

A series of ablation experiments were conducted to evaluate each module’s functionality within the model. The dataset selects the DIOR remote sensing dataset in Section 3.2 of the experiment, and performs training and testing under the same conditions. We utilized the YOLOv7 model as a foundation. In order to enhance the model’s performance, we introduced the BRA module (referred to as YOLOv7-b) and replaced the WIoUv3 loss function (referred to as YOLOv7-w). Additionally, we evaluated the combined effects of both modifications with our YOLOv7-bw variant, conducting ablation experiments accordingly.

The outcomes of these experiments are presented in Table 3, indicating that our YOLOv7-bw model achieved the most favorable results on the DIOR Dataset. Remarkably, it achieved a precision level of 85.63 when an IoU threshold of 0.5 was applied.

Figure 9 The (a) precision; (b) recall; (c) mAP@ 0.5; and (d) [email protected]:0.95 curves of YOLOv7 and YOLOv7-bw.

As shown in Table 4, we further conducted fine-grained ablation experiments on selected critical categories, including small dense targets (vehicles, ships, airplanes) and large targets (golffield, stadium). It can be observed that both WIoUv3 and BRA modules contribute positively across all categories. Specifically, the BRA module improves detection accuracy for small dense targets like vehicles and ships by better focusing attention on dense regions. Meanwhile, the WIoUv3 loss enhances bounding box regression, leading to improved localization. The combined YOLOv7-bw achieves the highest performance, particularly notable on small targets, while maintaining competitive results on large objects. This clearly demonstrates that the strategic integration of BRA, WIoUv3, and DCNv2 collectively contributes to the significant performance gains observed, validating the effectiveness of our innovative design choices.

Simultaneously, we graphed the accuracy measurements of precision, recall, mAP@ 0.5, and [email protected]:0.95 for both YOLOv7-bw and the original YOLOV7 source code, as exemplified in Figure 9. Among them, Precision is how many of the detected categories are accurate, and Recall is how many of all accurate categories have been detected. Besides, [email protected] and [email protected]:0.95 refer to the average value of all categories of AP under different IOU thresholds, which is the most important indicator of detection accuracy. According to the data presented in Figure 9, we can see that our YOLOv7-bw has a faster convergence speed, and it is close to being stable at around 50 rounds. Furthermore, YOLOv7 achieved an overall yield of 86.93%, 79.38%, 83.70%, and 63.9% in its final iteration. Conversely, our YOLOv7-bw exhibited a superior performance with respective scores of 90.73%, 80.13%, 85.63%, and 65.93%. These results unequivocally establish that YOLOv7-bw surpasses YOLOv7 in precision, recall, [email protected], and [email protected]:0.95. The inherent features of the remote sensing image encompass the qualities of petite targets, particularly when frequently exposed to atmospheric conditions, resulting in targets being obscured in shadows, thereby significantly amplifying the challenge of detection. This drawback restricts its progress in the realm of food provision. Figure 10 provides a visual representation of this scenario. In (a), the outcomes of the YOLOv7 assessment are presented, wherein three vehicles went undetected. However, upon incorporating WIoUv3 to YOLOv7, despite the continued omission of one vehicle due to its distinct dynamic non-monotonic focal mechanism, it can be observed in (b) that both the confidence level of object detection and the blurriness within the shadow have been enhanced to a certain degree. Consequently, the detection capabilities for all diminutive targets have been improved. Moreover, the heat map generated by the model pays less attention to noise such as trees. To further quantify these observations, we computed MAS within ground-truth bounding boxes. The results show that YOLOv7-bw achieves an MAS of 0.74, compared to 0.59 for YOLOv7, indicating a more refined and concentrated attention mechanism that effectively suppresses background distractions while enhancing target detection. Another noteworthy characteristic of remote sensing images is the concentration of small targets in specific areas. The BRA module shifts its focus towards these regions, employing a precise attention mechanism that aligns with the unique attributes of remote sensing imagery. This reality is depicted in Figure 11. The left image (a) exemplifies the detection outcome achieved by running the original YOLOv7 source code, while the right image (b) showcases the detection outcome after integrating the BRA module into the original YOLOv7 source code. Adding the BRA module increases the number of cars detected in the dense cluster, reducing false identifications of chimneys on the roof as cars. Additionally, the heat map reveals that the model pays less attention to the chimneys on the roof after implementing the BRA module. A quantitative comparison further confirms this improvement—after incorporating the BRA module, MAS within target regions rose from 0.67 (YOLOv7) to 0.81 (YOLOv7-bw), indicating a more refined ability to concentrate on relevant objects while effectively filtering out background noise. Overall, with the inclusion of BRA and the replacement of WIoUv3, YOLOv7 has significantly improved its detection capabilities for small, indistinct and dense targets. This enhancement enables its more effective application in agricultural scenarios, thereby contributing to facilitating food supply.

Compared to conventional detectors such as Faster R-CNN and SSD, which primarily rely on fixed receptive fields and anchor-based mechanisms, YOLOv7-bw demonstrates superior performance in detecting small and densely packed targets. This improvement can be attributed to the integration of the Bi-level Routing Attention (BRA) module, which dynamically adjusts attention to dense object regions, unlike traditional static attention mechanisms. Additionally, the adoption of DCNv2 enables the network to adapt to object deformations and spatial variance, a limitation observed in earlier architectures like RetinaNet and standard YOLOv7. Moreover, by incorporating WIoUv3 as the loss function, YOLOv7-bw improves bounding box regression precision for small targets, addressing the localization issues prevalent in previous methods. These enhancements collectively result in better feature representation and localization accuracy, as evidenced by the higher mAP and recall values. Our results align with previous studies emphasizing the importance of attention mechanisms and deformable convolutions in complex scenarios, while our model further refines these strategies to suit remote sensing imagery.

Figure 10 The influence of WIoUv3 to YOLOv7 (a) the outcomes of YOLOv7 in which three vehicles were undetected; (b) the outcomes of YOLOv7 implementing WIoUv3 that both the confidence level of object detection and the blurriness within the shadow have been enhanced to a certain degree, ameliorating the detection capabilities for all diminutive targets.

Figure 11 The influence of BRA on YOLOv7 (a) the outcomes of YOLOv7; (b) the outcomes of YOLOv7 implementing BRA; The first line is the detection results and the second line is the heat maps.

3.5 Analysis of food supply via DIOR Dataset

Remote sensing technology is widely applied across various fields in modern society, including monitoring transportation such as vehicles, ships, and airplanes. These monitoring activities contribute to traffic management and safety and provide robust support for optimizing food supply. Real-time monitoring of this transportation allows for a better understanding of road congestion, enabling the implementation of corresponding measures to improve traffic flow. This is crucial for the food supply chain as traffic congestion can lead to delays in food transportation, affecting the freshness and quality of food. On the other hand, by timely understanding the traffic situation, managers can more accurately predict the arrival time of food, thereby optimizing inventory management. This helps to prevent inventory backlog or shortages, increase inventory turnover, and reduce inventory costs. In the DIOR remote sensing image Dataset, the proportion of vehicles, ships, and airplanes exceeds 50% of the total, enhancing the detection accuracy for the DIOR Dataset, which is essentially an improvement in the efficiency of monitoring transportation. Simultaneously, the improved accuracy in detecting other categories in the Dataset further validates the robustness of our algorithm. Finally, to further demonstrate the critical role of our proposed method in food supply applications, we have selected 8 images for presentation, primarily focusing on vehicles. The actual results are shown in Figure 12. Our algorithm has detected almost all transportation in the picture, which plays a key role in logistics monitoring in food supply applications. However, there are still some small shortcomings. For example, in picture a, there is a missed detection of a vehicle, and in picture b, there is an incorrect detection of an airplane. In the next research, we will further solve these problems. We hope our algorithm can be further applied in the field of food supply.

Figure 12 Visualization of multiclass object-detection results using the proposed YOLOv7-bw on the DIOR Dataset. (a) ships and vehicles; (b) ships; (c) airplanes; (d) airplanes; (e) - (h) vehicles.