Lightweight remote sensing target detection method and system

Abstract

The invention discloses a lightweight remote sensing target detection method and system, and relates to the technical field of target detection, and the method comprises the steps: obtaining a to-be-detected target remote sensing image; and inputting the target remote sensing image into the trained improved YOLOv8n network model, and performing target detection on the target remote sensing image through the improved YOLOv8n network model to obtain a target detection result. According to the method, the lightweight MobileNetV3 network is adopted as a backbone network, and the reasoning speed is increased by using a deep separable convolution and efficient residual connection structure; through a nonlinear activation function and linear projection in residual connection, the feature expression capability is improved; a multi-scale attention expansion mechanism is introduced in front of a detection head, the receptive field of the model is expanded while the calculation cost is reduced and the processing efficiency is improved, and finally, a DWR module is introduced, feature fusion is optimized, and the detection capability of small targets and shielded targets in remote sensing images is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of target detection, and more particularly to a lightweight remote sensing target detection method and system. Background Art

[0002] Remote sensing technology is an important means of obtaining information about the Earth's surface and has been widely used in fields such as military reconnaissance, disaster assessment, resource surveys, and environmental monitoring. The spatial resolution of images continues to increase with advances in remote sensing technology, making it possible to extract target information from remote sensing images. Remote sensing target detection achieves precise location and identification of specific targets through the precise analysis and processing of remote sensing images. Small target detection in remote sensing images faces the following major challenges: 1. Limitations of feature extraction and selection: Early remote sensing detection methods relied on manual feature extraction and selection, which limited model generalization and automation, often requiring extensive expertise and human intervention. However, traditional CNN approaches are limited by perceptual constraints when processing high-resolution remote sensing images, resulting in poor detail capture and efficiency when extracting features. 2. Complexity and diversity of remote sensing scenes: A major challenge facing remote sensing target detection is the inherent complexity and diversity of remote sensing scenes. Different geographic features, object types, and environmental conditions result in extremely high variability in image data, which poses a challenge to target detection algorithms. 3. Efficiency and memory challenges of deep learning models: With the application of deep learning in remote sensing detection, the computational efficiency and memory usage of the model have become new issues.

[0003] Therefore, how to provide a lightweight remote sensing target detection method and system to efficiently and accurately detect target objects from a large number of remote sensing images is an urgent problem that needs to be solved by those skilled in the art. Summary of the Invention

[0004] In view of this, the present invention provides a lightweight remote sensing target detection method and system, which aims to enhance the target feature extraction capability in remote sensing images and improve model performance and processing efficiency by introducing lightweight network structure, feature fusion strategy and adding attention mechanism and other technical means.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: a lightweight remote sensing target detection method, comprising:

[0006] Acquire a remote sensing image of the target to be detected;

[0007] Inputting the target remote sensing image into the trained improved YOLOv8n network model, performing target detection on the target remote sensing image through the improved YOLOv8n network model to obtain a target detection result;

[0008] Among them, the improved YOLOv8n network model includes: using lightweight MobileNetV3 as the backbone network for feature extraction, introducing the DWR module in the C2f of the Neck network for feature fusion, and introducing a multi-scale expansion attention mechanism before the detection head to capture multi-scale contextual information.

[0009] Preferably, the lightweight MobileNetV3 introduces an inverted residual structure and a linear bottleneck layer to optimize the bneck module; wherein, the bneck module performs feature extraction through depth-separable convolution and SE channel attention mechanism.

[0010] Preferably, the DWR module is designed in a residual manner, draws multi-scale context information inside the residual, and fuses the feature maps generated by the multi-scale receptive fields.

[0011] Preferably, multi-scale context information is drawn inside the residual, and the feature map generated by the multi-scale receptive field is fused, including:

[0012] Regional residualization: 3×3 convolution is used for initial feature extraction, and then combined with batch normalization layer and RELU layer to generate regional residual features of different region sizes;

[0013] Semantic residualization: group the regional residual features, set dilated convolution with different dilation rates to perform morphological filtering on the grouped regional residual features;

[0014] After regional residualization and semantic residualization, the processed feature maps are fused.

[0015] Preferably, the DWR module is introduced into C2f of the Neck network, including:

[0016] The DWR module is used to replace the second ordinary convolution in Bottleneck in C2f, and the activation function adopts the GELU function.

[0017] Preferably, a multi-scale dilated attention mechanism is introduced before the detection head to capture multi-scale contextual information, including:

[0018] Linearly project the feature map X to obtain the query Q, key K and value V;

[0019] The channels of the feature map are divided into multiple different heads for parallel processing, and the number of holes in each head corresponds to a different receptive field;

[0020] A self-attention operation is performed on the window around the query block, and then the features of different heads are concatenated and fed into the linear layer.

[0021] Preferably, the expression of the multi-scale expansion attention mechanism is as follows:

[0022] h i =SWDA(Q i ,K i ,V i ,r i ),1≤i≤n;

[0023] Among them, Q i ,K i ,V i represents the feature map slice fed to the i-th head, r i is the expansion rate of the i-th head, h i represents the output of the i-th attention head;

[0024] Output Connect them together and send them to the linear layer for feature aggregation. The result after linear aggregation is expressed as:

[0025] X=Linear(Concat[h1,...,h n ]).

[0026] Preferably, a lightweight remote sensing target detection system comprises: an image acquisition module for acquiring a remote sensing image of a target to be detected;

[0027] The target detection module is used to input the target remote sensing image into the trained improved YOLOv8n network model, perform target detection on the target remote sensing image through the improved YOLOv8n network model, and obtain a target detection result.

[0028] As can be seen from the above technical solution, compared with the prior art, the present invention discloses a lightweight remote sensing target detection method and system, including: obtaining a remote sensing image of a target to be detected; inputting the target remote sensing image into a trained improved YOLOv8n network model, and performing target detection on the target remote sensing image using the improved YOLOv8n network model to obtain a target detection result. The improved YOLOv8n network model includes: using a lightweight MobileNetV3 as the backbone network for feature extraction, introducing a DWR module in the C2f of the Neck network for feature fusion, and introducing a multi-scale dilated attention mechanism before the detection head to capture multi-scale contextual information. This paper adopts the lightweight MobileNetV3 network to replace the backbone network of YOLOv8, and uses deep separable convolution and efficient residual connection structure to improve the inference speed; improves the feature expression ability through nonlinear activation function and linear projection in residual connection; introduces a multi-scale dilated attention mechanism before the detection head, which reduces the computational cost and improves the processing efficiency while expanding the receptive field of the model, thereby enhancing the model's recognition accuracy for targets of different sizes and details, especially small targets; finally, introduces the DWR module into the C2f of the Neck network of YOLOv8 to optimize feature fusion, significantly improving the detection ability of small targets and occluded targets in remote sensing images. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are merely embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying any creative work.

[0030] Figure 1 This is a diagram of the MDM-YOLO structure provided by an embodiment of the present invention.

[0031] Figure 2 This is a diagram of the MobileNetV3 architecture provided by an embodiment of the present invention.

[0032] Figure 3 A schematic diagram of a bneck module provided in an embodiment of the present invention.

[0033] Figure 4 A schematic diagram of a DWR module provided in an embodiment of the present invention.

[0034] FIG5( a ) is a schematic diagram of a C2f module provided in an embodiment of the present invention.

[0035] FIG5( b ) is a schematic diagram of a C2f_DWR module provided in an embodiment of the present invention.

[0036] Figure 6 A schematic diagram of the MSDA module provided in an embodiment of the present invention.

[0037] Figure 7 A comparison chart of MDM-YOLO detection effects in various scenarios provided by the embodiments of the present invention.

[0038] Figure 8 Schematic diagram of a remote sensing target detection method based on improved YOLOv8n provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0039] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0040] The embodiment of the present invention discloses a lightweight remote sensing target detection method, such as Figure 8 As shown, including:

[0041] Acquire a remote sensing image of the target to be detected;

[0042] Inputting the target remote sensing image into the trained improved YOLOv8n network model, performing target detection on the target remote sensing image through the improved YOLOv8n network model to obtain a target detection result;

[0043] Among them, the improved YOLOv8n network model includes: using lightweight MobileNetV3 as the backbone network for feature extraction, introducing the DWR module in the C2f of the Neck network for feature fusion, and introducing a multi-scale expansion attention mechanism before the detection head to capture multi-scale contextual information.

[0044] Specifically, the embodiment of the present invention improves the YOLOv8n network model: the detailed architecture diagram of MDM-YOLO (MobileNetV3-DWR-MSDA-YOLO) is as follows Figure 1As shown in the figure, MDM-YOLO is an attempt to improve the object detection performance of YOLOv8n in the field of remote sensing image recognition. Compared with the original YOLOv8, the improvements selected in this embodiment mainly include three aspects: backbone network optimization, algorithm module, and the introduction of attention mechanism. These correspond to three applied technologies: MobileNetV3, Multi-Scale Dilated Attention (MSDA), and DWR (Dilation-wise Residual Module).

[0045] This model replaces the YOLOv8 backbone with a MobileNetV3 small network, introduces a multi-scale dilated attention mechanism before the detection head, and incorporates a DWR module into the C2f of the Neck network. These improvements not only reduce computing resource requirements, enhance feature extraction and fusion, but also improve the model's generalization and robustness. Specifically:

[0046] (1) Reduced computing resource requirements: Lightweight network design reduces the number of parameters and computational complexity.

[0047] (2) Expanded model receptive field: The multi-scale expanded attention mechanism expands the receptive field without increasing the amount of computation, thereby enhancing the model's recognition accuracy for targets of different sizes and details.

[0048] (3) Improved feature fusion effect: The DWR module enhances the fusion of deep and shallow features.

[0049] (4) Balance between real-time performance and accuracy: Reduce latency while maintaining high accuracy.

[0050] (5) Enhanced robustness: better cope with complex scenarios and noise interference.

[0051] Specifically, the lightweight MobileNetV3 introduces an inverted residual structure and a linear bottleneck layer to optimize the bneck module; wherein, the bneck module performs feature extraction through depth-wise separable convolution and SE channel attention mechanism.

[0052] Specifically, lightweight model design is crucial to meet the deployment requirements of remote sensing target detection tasks on mobile devices. The MobileNetV3 series, available in two versions, small and large, focuses on reducing model size and computational load while maintaining high accuracy. A major improvement in MobileNetV3 lies in its use of an inverted residual structure and separable convolutions, combined with the SE channel attention mechanism and the Hard Sigmoid nonlinear activation function to enhance feature representation capabilities.

[0053] MobileNetV3 uses Neural Architecture Search (NAS) to build the network, optimizes the bneck module structure, and introduces an inverted residual structure and a linear bottleneck layer to improve detection efficiency. Its core bneck module uses depthwise separable convolution and the SE channel attention mechanism to achieve efficient feature extraction, thereby improving the network's lightweightness.

[0054] Figure 2 and Figure 3 The overall architecture of MobileNetV3 and the detailed structure of the bneck module are shown respectively. Figure 2 As shown in the figure, MobileNetV3 uses Neural Architecture Search (NAS) to build the network, optimizes the bneck module structure, and introduces an inverted residual structure and a linear bottleneck layer to improve detection efficiency. Its core bneck module uses depthwise separable convolution and the SE channel attention mechanism to achieve efficient feature extraction, thereby improving the network's lightweightness.

[0055] like Figure 3 As shown in the figure, the bneck module first uses a 1x1 convolutional layer to expand the dimension of the input features, then feeds them into a 3x3 depthwise separable convolutional layer, keeping the number of channels unchanged. Next, the features pass through a global average pooling layer and are weighted using the SE channel attention mechanism. Finally, a 1x1 convolutional layer is used to reduce the dimensionality and fuse shallow feature information to achieve feature reuse.

[0056] The lightweight MobileNetV3, used as the backbone for YOLOv8 in remote sensing detection tasks, offers the following benefits: In terms of deep and shallow feature extraction, optimized modules and global average pooling layers significantly improve feature extraction capabilities, capturing both local and global image features and enhancing detection accuracy and robustness. Furthermore, its lightweight design and deep optimization strategies significantly increase inference speed, making it suitable for real-time applications. This addresses, to some extent, the shortcomings of YOLOv8n in terms of model lightweightness and feature extraction capabilities.

[0057] Specifically, the DWR module is designed in a residual manner, draws multi-scale context information inside the residual, and fuses the feature maps generated by the multi-scale receptive fields.

[0058] Specifically, we draw multi-scale context information inside the residual and fuse the feature maps generated by multi-scale receptive fields, including:

[0059] Regional residualization: The initial feature extraction uses 3×3 convolution, and then combines the batch normalization layer and the RELU layer to generate a series of regional residual features RR of different region sizes;

[0060] Semantic residualization: The regional residual features RR are grouped, and dilated convolutions with different dilation rates are set to perform morphological filtering on the grouped regional residual features;

[0061] After regional residualization and semantic residualization, the processed feature maps are fused.

[0062] Specifically, in YOLOv8, Backbone is responsible for feature extraction, while Neck is responsible for feature fusion. For remote sensing detection tasks, in traditional convolutional neural networks, due to the limitation of the receptive field, it is difficult for the network to capture the features of large-scale and small-scale targets at the same time. Secondly, background information is often mixed with target information. Therefore, the effect of deep and shallow feature fusion in Neck directly affects the performance of the final target detection. A good fusion strategy can not only enhance feature representation, optimize computing resources, and promote gradient transfer, but also enhance the network's generalization ability and ability to cope with multi-scale targets. Therefore, the embodiment of the present invention improves the C2f structure in Neck, adds the DWR module, and utilizes its advantages of multi-scale feature extraction, efficient feature fusion, improved network adaptability and optimized resource utilization to significantly improve the effect of deep and shallow feature fusion.

[0063] Feature maps at different levels have different receptive fields and semantic strengths. The DWR module allows for dynamic adjustment of the expansion rate according to different task requirements and data characteristics. This flexibility ensures the effectiveness of the module in remote sensing application scenarios. The DWR module is designed in a residual manner, such as Figure 4 As shown in the figure. Within the residual, a two-step method is used to effectively draw multi-scale contextual information, and then the feature maps generated by the multi-scale receptive fields are fused. The first step: preliminary feature extraction uses a 3×3 convolution, combined with a batch normalization layer and a RELU layer to generate a series of regional residual features RR of different region sizes; the second step: semantic residualization divides the RR features into three groups, and sets three branches of dilated convolution with dilation rates of 1, 3, and 5 for morphological filtering, and the output channel capacity of the first branch is expanded to twice that of the other branches. Finally, after regional residualization and semantic residualization, the DWR module merges or fuses the processed feature maps, which are the spliced ​​and integrated multi-scale contextual feature maps.

[0064] The embodiment of the present invention uses C2f_DWR to replace C2f in Neck. The schematic diagrams of the C2f and C2f_DWR module structures are shown in Figures 5(a) and 5(b). In Figure 5(a), input segmentation is to convert the input feature map into twice the number of hidden channels through the first convolution layer, and then split it into two parts according to the channel, one of which is directly used as the residual path, and the other part enters the subsequent processing. All processed features (including the residual path and Bottleneck output) are spliced ​​in the channel dimension and integrated into the output channel number through the second convolution layer. In Figure 5(b), DWR is used to replace the second ordinary convolution in Bottleneck in C2f, and the activation function is replaced by GELU from SiLU. The DWR module combines the two steps of regional residual and semantic residual to improve the expressiveness of the feature map, which directly enhances the fusion ability of deep and shallow layer features. In addition, compared with the SiLU activation function, GELU provides a non-zero gradient in the negative value range, effectively alleviating the gradient disappearance problem and improving the stability of network training. At the same time, it can accelerate the convergence of the model, thereby improving the overall training efficiency. The combination of the DWR module and the GELU activation function improves the model's adaptability to different types of data sets.

[0065] Specifically, the DWR module is introduced into C2f of the Neck network, including:

[0066] The DWR module is used to replace the second ordinary convolution in Bottleneck in C2f, and the activation function adopts the GELU function.

[0067] Specifically, a multi-scale dilated attention mechanism is introduced before the detection head to capture multi-scale contextual information, including:

[0068] Linearly project the feature map X to obtain the query Q, key K and value V;

[0069] The channels of the feature map are divided into multiple different heads for parallel processing, and the number of holes in each head corresponds to a different receptive field;

[0070] A self-attention operation is performed on the window around the query block, and then the features of different heads are concatenated and fed into the linear layer.

[0071] Specifically, for remote sensing detection tasks, a larger receptive field is essential for achieving high-precision target detection in diverse and complex remote sensing scenarios, as it helps the model capture sufficient contextual information. There are various ways to increase the receptive field, including dilated convolution, SPP, increasing the convolution kernel size, and adding more network layers. Dilated convolution, among others, has the advantages of increasing the receptive field without downsampling, increasing the receptive field with the same convolution kernel size without increasing computational effort, and dynamically adjusting the dilation ratio to control the receptive field. Therefore, it is widely used to increase the receptive field of remote sensing detection models.

[0072] After stacking 3×3 ordinary convolutions, the receptive field can achieve the effect of 5×5 and 7×7 ordinary convolutions. However, the dilated convolution can increase the receptive field without increasing the number of parameters (parameters = convolution kernel weight + bias). The receptive field is calculated as follows:

[0073] k'=k+(k-1)×(r-1);

[0074] RF i+1 =RF i +(k i '-1)×S i ;

[0075]

[0076] Among them, k represents the convolution kernel size, r is a hyperparameter representing the number of holes, k' represents the equivalent convolution kernel size, and the receptive field of the i-th layer is RF i , S i It represents the stride product of the i-th layer and all previous layers, Stride j represents the step size of the j-th layer.

[0077] As can be seen from the formula, the stride of the i+1 layer is i+1 It does not affect the receptive field of the i+1th layer, and the receptive field is independent of Padding. However, since the calculation method of the dilated convolution is similar to the checkerboard format, that is, r-1 zeros are filled between the convolution kernels of the current layer, which come from an independent set that has no mutual dependence in the previous layer, there is a lack of correlation between the convolution results of this layer, which can easily lead to local information loss. In remote sensing image processing, each pixel may contain important ground object information, such as small vehicles, buildings, etc. Therefore, the embodiment of the present invention chooses to add Multi-Scale DilatedAttention after feature fusion. This mechanism captures multi-scale contextual information by expanding the receptive field through dilated convolution, and combines the local attention mechanism to reduce the loss of local information.

[0078] MSDA Figure 6 As shown, given a feature map X, a linear projection is performed to obtain the corresponding query (Q), key (K), and value (V). The feature map's channels are then split into four different heads for parallel processing, with dilation numbers r = 1, 2, 3, and 4. The sizes of the participating receptive fields of the corresponding heads are 3×3, 5×5, 7×7, and 9×9, respectively. Self-attention is performed between the colored blocks in the window surrounding the red query block. The features of the different heads are then concatenated and fed into a linear layer. This embodiment uses four parallel heads to enable MSDA to cover a wider receptive field, making it more suitable for remote sensing image processing.

[0079] By setting different dilation rates for different heads, MSDA effectively aggregates semantic information at different scales. The model can perceive contextual information in a wider area and capture image features at different scales, thereby improving the model's detection capability without increasing complex operations and additional computational costs.

[0080] Specifically, the expression of the multi-scale expansion attention mechanism is as follows:

[0081] h i =SWDA(Q i ,K i ,V i ,r i ),1≤i≤n;

[0082] Among them, Q i ,K i ,V i represents the feature map slice fed to the i-th head, r i is the expansion rate of the i-th head, h i Represents the output of the i-th attention head (Head), each head corresponds to a different expansion rate r i and scale s i .

[0083] Output Connect them together and send them to the linear layer for feature aggregation. The result after linear aggregation is expressed as:

[0084] X=Linear(Concat[h1,...,h n ]).

[0085] In a specific embodiment of the present invention, a lightweight remote sensing target detection system includes: an image acquisition module for acquiring a remote sensing image of a target to be detected;

[0086] The target detection module is used to input the target remote sensing image into the trained improved YOLOv8n network model, perform target detection on the target remote sensing image through the improved YOLOv8n network model, and obtain a target detection result.

[0087] The embodiment of the present invention replaces the backbone network of YOLOv8 with a lightweight MobileNetV3-small network, and utilizes depthwise separable convolution and efficient residual connection structures to improve inference speed. The feature expression capability is improved through nonlinear activation functions and linear projections in residual connections. A multi-scale dilated attention mechanism (MSDA) is introduced before the detection head to reduce computational costs, improve processing efficiency, and expand the receptive field of the model, thereby enhancing the model's recognition accuracy for targets of different sizes and details, especially small targets. Finally, a DWR (Dilation-wise Residual Module) module is introduced into the C2f of the Neck network of YOLOv8 to optimize feature fusion and significantly improve the detection capability of small and occluded targets in remote sensing images.

[0088] Experimental results and analysis

[0089] In the field of remote sensing, the DIOR dataset and the NWPU VHR-10 dataset are two commonly used standard datasets. This paper presents ablation and comparative experiments on the DIOR dataset, and generalization experiments on the NWPU VHR-10 dataset. It then presents comparative experiments with different MDM-YOLO improvements, ablation experiments, and other existing models. Finally, it analyzes the detection and real-time performance of MDM-YOLO on the DIOR dataset.

[0090] Comparison of different improvements, ablation experiments, and results analysis

[0091] In the comparative experiments on different backbone network improvements, Table 2 shows that although LSKNet (Large Selective Kernel Network) can also extract global contextual information by dynamically adjusting the spatial receptive field, prematurely expanding the receptive field usually requires more convolutional layers or larger convolution kernels, which directly leads to an increase in the number of model parameters and floating-point operations. This not only means that the model requires more storage space, but also increases the computational burden during model training and inference. At the same time, prematurely expanding the receptive field leads to increased network depth, exacerbating the gradient vanishing and explosion problems, and ultimately resulting in detection accuracy that is inferior to the baseline model. The recurrent structure of RecCoIV1 (Recurrent Convolutional Instance Vector 1) requires high computing resources. In addition, the repeated block structure of RecCoIV1 is too complex for remote sensing detection tasks and fails to effectively capture the characteristics of remote sensing data. The detection accuracy is also lower than the baseline. Next, we conducted comparative tests on three versions of the MobileNet series. The experimental results are shown in Table 1 below. The three versions achieved improvements of 1.3, 1.1, and 1.4, respectively, in mAP@0.5% and 1.6, 1.4, and 1.7, respectively, over the baseline model, while maintaining a minimal increase in the number of parameters and floating-point operations. This embodiment of the present invention ultimately selected MobileNetV3 to replace the YOLOv8 backbone network.

[0092] Table 1 Comparative experiments of various backbone network improvements and MobileNetV3

[0093]

[0094] Table 2 shows the experimental results of experiments with various attention mechanism improvements. Large Separable Kernel Attention (LSKA) enriches feature representations by using kernels of varying sizes, improving the model's focus on important features in the input data. With a slight increase in the number of parameters and floating-point operations, mAP@0.5% and mAP@0.5:0.95% improved by 0.9 and 1.7, respectively. ACmix_Attention enhances the performance of convolutional neural networks by fusing features from different layers, resulting in improved accuracy. MSDA uses different dilation rates to capture multi-scale local contextual information and uses a local cross-attention mechanism to better capture subtle features of small objects. This helps the network better understand multi-scale structures and details in images, improving model detection performance. mAP@0.5% improves by 1.3, and mAP@0.5:0.95% improves by 1.8, compared to the baseline, with a slight increase in parameters and floating-point operations. EMA smoothes model parameters by weighted averaging, improving mAP@0.5% and mAP@0.5:0.95% by 0.4 and 0.5, respectively, though this performance is lower than that of other attention mechanisms. TripletAttention enhances the performance of convolutional neural networks by combining information from three different feature maps, strengthening the integration of local and global information and dynamically weighting feature channels, resulting in improved accuracy. In summary, the introduction of the MSDA mechanism can further improve model detection performance and robustness, making it more suitable for remote sensing target detection tasks.

[0095] Table 2 Comparative experiments with different attention mechanisms

[0096]

[0097] As shown in Table 3 below, in ablation experiments on the DIOR dataset, replacing the backbone network with MobileNetV3 significantly improves the model's object detection accuracy, increasing mAP@0.5% by 1.4 and mAP@0.5:0.95% by 1.7 compared to the baseline model. This demonstrates the effectiveness of the MobileNetV3 lightweight network in capturing small objects, but also slightly increases the model's parameters and computational overhead. Introducing the DWR module in C2f also significantly improves accuracy, demonstrating the model's ability to effectively transfer features. Integrating the MSDA mechanism improves mAP@0.5% by 1.3 and mAP@0.5:0.95% by 1.8. The experimental results show that combining all three improvements achieves the highest accuracy, with mAP@0.5% and mAP@0.5:0.95% reaching 87.2 and 65.6, respectively. Despite the increase in parameters and GFLOPs, the model still meets real-time requirements. In summary, the model is well-suited for remote sensing detection tasks.

[0098] Table 3 Ablation comparison experiments on the DIOR dataset

[0099]

[0100] Comparative analysis of the DIOR dataset and generalization verification of the NWPU VHR-10 dataset

[0101] To evaluate the performance of the model, the present invention conducted comparative experiments on the DIOR dataset. The experimental results are shown in Table 4. The present invention uses several different models for comparison, including traditional two-stage object detection models such as Faster-RCNN; one-stage models such as SSD, RetinaNet, CenterNet, and YOLO series; and some of the latest deep learning methods such as RT-DETR. The experimental results in Table 4 below show that most one-stage detectors are faster than two-stage and state-of-the-art algorithms. Two-stage detectors have a high number of parameters and floating-point operations, which do not meet the requirements for lightweight and real-time remote sensing detection models. Among single-stage detectors, the YOLO family performs best, achieving high detection accuracy while maintaining low parameter count and GFLOPs. Its FPS metric outperforms other models, better meeting the real-time requirements. By comparison, the model proposed in this embodiment of the present invention performs well in the experiments. Although its FPS is 385, slower than YOLOv8n, it achieves 87.2% mAP@0.5%, ranking first and achieving improvements of 19.0, 24.6, 15.8, and 16.4 over Faster-RCNN, SSD, RetinaNet, and CenterNet, respectively. These experimental results fully demonstrate that the improved model proposed in this embodiment of the present invention has excellent detection capabilities and meets the requirements for real-time and lightweight models. In summary, the proposed model is suitable for remote sensing image object detection tasks.

[0102] Table 4 Comparative experimental results of different models on the DIOR dataset

[0103]

[0104]

[0105] To evaluate the model's performance and verify its generalization, the present invention conducted validation experiments on an expanded dataset of the public NWPU VHR-10 dataset. The experimental results are shown in Table 5. Compared with the YOLOv8n algorithm, MDM-YOLAO achieved a 1.1% improvement in mAP@0.5% and a 0.7% improvement in mAP@0.5:0.95%, effectively reducing missed and false positives. This demonstrates that the improved model presented in the present invention has good generalization performance on remote sensing image datasets and is suitable for remote sensing monitoring tasks.

[0106] Table 5 Comparative experimental results on the NWPU VHR-10 dataset before and after model improvement

[0107]

[0108] Detection effect analysis

[0109] In order to intuitively demonstrate the detection effect of the improved algorithm, a comparative test was conducted on four scenarios of the DIOR dataset using YOLOv8n and the MDM-YOLO algorithm of the embodiment of the present invention. The results are as follows: Figure 7 As shown in the figure, the left side is the original image, the middle is the YOLOv8n detection result, and the right side is the MDM-YOLO detection result. Figure 7 As shown in (a) in the figure, YOLOv8n detects medium-sized targets more accurately, but misses and misdetects small targets such as cars and blue boxes, while MDM-YOLO can detect these small targets more accurately. Figure 7 As shown in (b), YOLOv8n fails to detect the small playground in the upper right corner, while MDM-YOLO fully detects it with high accuracy. Figure 7 As shown in (c) in Figure 2, MDM-YOLO has a significant improvement over YOLOv8n in detecting ports and ships. Figure 7 As shown in (d), MDM-YOLO can better identify the tennis court obscured by leaves, while YOLOv8n misses the detection. In summary, MDM-YOLO performs better than YOLOv8n in small target detection, dense scenes, complex backgrounds, and occlusion scenes.

[0110] This embodiment of the present invention proposes a lightweight remote sensing target detection method and system. This method replaces the YOLOv8 backbone network with a lightweight MobileNetV3-small network, adds an MSDA attention mechanism before each detection head, and introduces DWR into the C2f of the Neck network. Compared with the original YOLOv8n model, MDM-YOLO surpasses it in all aspects. MDM-YOLO not only improves the detection accuracy of targets, especially small targets, but also meets the requirements of a lightweight model. On the DIOR dataset, MDM-YOLO generally outperforms various target detection algorithms in terms of accuracy and real-time performance.

[0111] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Reference can be made to the common and similar parts between the various embodiments. For the devices disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple, and the relevant parts can be referred to the method description.

[0112] The above description of the disclosed embodiments is intended to enable one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined in the embodiments of the present invention may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention is not limited to the embodiments shown in the examples of the present invention, but is intended to conform to the widest scope consistent with the principles and novel features disclosed in the embodiments of the present invention.

Claims

1. A lightweight remote sensing target detection method, characterized in that: include: Acquire a remote sensing image of the target to be detected; Inputting the target remote sensing image into the trained improved YOLOv8n network model, performing target detection on the target remote sensing image through the improved YOLOv8n network model to obtain a target detection result; Among them, the improved YOLOv8n network model includes: using lightweight MobileNetV3 as the backbone network for feature extraction, introducing the DWR module in the C2f of the Neck network for feature fusion, and introducing a multi-scale expansion attention mechanism before the detection head to capture multi-scale contextual information.

2. A lightweight remote sensing target detection method according to claim 1, characterized in that: The lightweight MobileNetV3 introduces an inverted residual structure and a linear bottleneck layer to optimize the bneck module; wherein, the bneck module performs feature extraction through depthwise separable convolution and SE channel attention mechanism.

3. A lightweight remote sensing target detection method according to claim 1, characterized in that: The DWR module is designed in a residual manner, draws multi-scale context information inside the residual, and fuses the feature maps generated by multi-scale receptive fields.

4. A lightweight remote sensing target detection method according to claim 3, characterized in that: Draw multi-scale context information inside the residual and fuse the feature maps generated by multi-scale receptive fields, including: Regional residualization: 3×3 convolution is used for initial feature extraction, and then combined with batch normalization layer and RELU layer to generate regional residual features of different region sizes; Semantic residualization: group the regional residual features, set dilated convolution with different dilation rates to perform morphological filtering on the grouped regional residual features; After regional residualization and semantic residualization, the processed feature maps are fused.

5. A lightweight remote sensing target detection method according to claim 1, characterized in that: Introducing the DWR module into C2f of the Neck network, including: The DWR module is used to replace the second ordinary convolution in Bottleneck in C2f, and the activation function adopts the GELU function.

6. A lightweight remote sensing target detection method according to claim 1, characterized in that: A multi-scale dilated attention mechanism is introduced before the detection head to capture multi-scale contextual information, including: Linearly project the feature map X to obtain the query Q, key K and value V; The channels of the feature map are divided into multiple different heads for parallel processing, and the number of holes in each head corresponds to a different receptive field; A self-attention operation is performed on the window around the query block, and then the features of different heads are concatenated and fed into the linear layer.

7. A lightweight remote sensing target detection method according to claim 6, characterized in that: The expression of the multi-scale expansion attention mechanism is as follows: h i =SWDA(Q i ,K i ,V i ,r i ),1≤i≤n; Among them, Q i ,K i ,V i represents the feature map slice fed to the i-th head, r i is the expansion rate of the i-th head, h i represents the output of the i-th attention head; Output Connect them together and send them to the linear layer for feature aggregation. The result after linear aggregation is expressed as: X=Linear(Concat[h1,...,h n ])。 8. A lightweight remote sensing target detection system, applying a lightweight remote sensing target detection method according to any one of claims 1 to 7, characterized in that: include: An image acquisition module is used to acquire a remote sensing image of a target to be detected; The target detection module is used to input the target remote sensing image into the trained improved YOLOv8n network model, perform target detection on the target remote sensing image through the improved YOLOv8n network model, and obtain a target detection result.

Images