Lightweight remote sensing image target detection method and system suitable for edge device

Abstract

The invention relates to the technical field of computer vision, and provides a lightweight remote sensing image target detection method and system suitable for edge equipment. The method comprises the steps of 1, obtaining a remote sensing image data set and performing preprocessing to obtain a training set and a test set; 2, constructing a remote sensing target detection model; wherein the remote sensing target detection model takes a YOLOv8 network as a reference model, and introduces a feature selection aggregation down-sampling module, a heterogeneous receptive field feature enhancement module and a multi-branch multi-scale fusion attention mechanism module; step 3, performing end-to-end training on the remote sensing target detection model by using the training set; and 4, inputting the test set into a trained remote sensing target detection model, and outputting a target bounding box and a category prediction result. The method and the system provided by the invention can improve the detection precision of the multi-scale target in the remote sensing image and the actual deployment performance of the model.

Description

Technical Field

[0001] This invention relates to the field of computer vision technology, and in particular to a lightweight remote sensing image target detection method and system suitable for edge devices. Background Technology

[0002] With the rapid development of remote sensing imaging technology and high-resolution satellite imagery, target detection in remote sensing images is increasingly widely used in fields such as land resource surveys, urban management, and environmental monitoring. Remote sensing images often contain rich information about ground features, such as buildings, vehicles, aircraft, and ships. Target detection, as one of the key technologies in intelligent remote sensing image analysis, is of great significance for achieving automated information extraction. However, compared to natural images, optical remote sensing images (ORSI) have more complex backgrounds, varying scale differences, and weak semantic consistency, which significantly increase the technical challenges of target detection tasks.

[0003] Current remote sensing image target detection technology still faces many challenges. First, the background of remote sensing images is complex and varied, with interference elements such as buildings, vegetation, and shadows often being confused with foreground targets. This makes it difficult for detection models to accurately separate targets from the background without explicit guidance or auxiliary supervision, leading to a decrease in detection accuracy. Second, the significant differences between targets at multiple scales also place higher demands on detection capabilities, especially in small target detection tasks. Small target sizes and weak semantic representations often result in features being suppressed and boundaries becoming blurred, easily leading to missed detections and false detections. Furthermore, to improve detection performance, some methods introduce a large number of complex modules, such as multi-layered pyramid structures, attention mechanisms, and auxiliary branches. While these redundant structures improve accuracy, they significantly increase the computational burden and parameter size of the model, hindering real-time deployment on resource-constrained devices such as UAVs and satellite terminals. Finally, existing feature extraction methods still have limitations in expressive power. While high-level semantic features have good semantic abstraction capabilities, they often sacrifice spatial detail information, while low-level features, although possessing fine edge representation, lack semantic completeness. Existing fusion strategies cannot fully combine the advantages of both, restricting further improvements in overall detection performance. Summary of the Invention

[0004] To address the problems of insufficient target detection accuracy caused by complex background interference, inadequate small target detection capability, and redundant model structure in existing remote sensing image target detection methods, this invention proposes a lightweight remote sensing image target detection method and system suitable for edge devices, which can improve the detection accuracy of multi-scale targets in remote sensing images and the actual deployment performance of the model.

[0005] In a first aspect, the present invention provides a lightweight remote sensing image target detection method suitable for edge devices, comprising:

[0006] Step 1: Obtain the remote sensing image dataset and preprocess it to obtain the training set and test set;

[0007] Step 2: Construct a remote sensing target detection model; wherein the remote sensing target detection model uses the YOLOv8 network as the baseline model and introduces a feature selection aggregation downsampling module, a heterogeneous receptive field feature enhancement module, and a multi-branch multi-scale fusion attention mechanism module; the aggregation downsampling module is used to improve the expressive power of shallow features; the heterogeneous receptive field feature enhancement module is used to enhance the robustness and directional consistency of feature expression; the multi-branch multi-scale fusion attention mechanism module is used to improve the sensitivity and robustness of the remote sensing target detection model to targets in complex environments;

[0008] Step 3: Perform end-to-end training of the remote sensing target detection model using the training set;

[0009] Step 4: Input the test set into the trained remote sensing target detection model and output the target bounding box and category prediction results.

[0010] Further, in step 1, the preprocessing includes performing uniform size normalization on the remote sensing images in the remote sensing image dataset, then performing data augmentation using various online data augmentation strategies, and dividing the training set and test set according to a preset ratio; the various online data augmentation strategies include: Mosaic mosaicking, color perturbation, random cropping, and mirroring.

[0011] Furthermore, in step 2, the remote sensing target detection model uses the YOLOv8 network as the baseline model and introduces a feature selection aggregation downsampling module, a heterogeneous receptive field feature enhancement module, and a multi-branch multi-scale fusion attention mechanism module. Specifically, it includes replacing the CBS module and C2F module in the backbone network of the YOLOv8 network with the feature selection aggregation downsampling module and the heterogeneous receptive field feature enhancement module, respectively, and adding the multi-branch multi-scale fusion attention mechanism module at the connection between the neck network and the backbone network.

[0012] Furthermore, in step 2, the feature selection aggregation downsampling module is used to improve the expressive power of shallow features, and the specific processing flow is as follows:

[0013] The input features of the feature selection aggregation downsampling module are processed by a 3×3 convolution to obtain the first feature;

[0014] The first feature is input into the SPDConv module for processing to obtain the second feature;

[0015] The first feature is processed by the dilated convolutional block and the SPDConv module to obtain the third feature;

[0016] The second feature and the third feature are spliced ​​together by the splicing unit and then input into the average pooling block and the max pooling block respectively to obtain the fourth feature and the fifth feature;

[0017] The fourth and fifth features are processed by convolutional blocks and a sigmoid function to obtain attention weights.

[0018] The second feature and the third feature are weighted by channel multiplication using the attention weights respectively. The second feature and the third feature after channel multiplication weighting are then fused and processed by 1×1 convolution to obtain the output feature of the feature selection aggregation downsampling module.

[0019] Further, in step 2, the heterogeneous receptive field feature enhancement module is used to enhance the robustness and directional consistency of feature representation, specifically including: a 3×3 convolutional block, parallel directional convolutional branches, and a dynamic weight adjustment branch; wherein the directional convolutional branch includes a first branch and a second branch, the first branch including a horizontal convolutional block, a vertical convolutional block, and a 1×1 convolutional block connected in sequence, and the second branch including a vertical convolutional block, a horizontal convolutional block, and a 1×1 convolutional block connected in sequence; the dynamic weight adjustment branch includes a 3×3 convolutional block, an activation function layer, a 1×1 convolutional block, an activation function layer, a 1×1 convolutional block, a binary classification function, and a weight value layer connected in sequence;

[0020] The output of the heterogeneous receptive field feature enhancement module is obtained by weighted and fused with the outputs of the directional convolution branch and the dynamic weight adjustment branch, and then added to the input features of the heterogeneous receptive field feature enhancement module using a residual method.

[0021] Furthermore, in step 2, the multi-branch multi-scale fusion attention mechanism module is used to improve the sensitivity and robustness of the target detection model to targets in complex environments. Specifically, it includes a spatial attention branch and a multi-scale channel enhancement branch. The spatial attention branch includes a parallel average pooling layer and a max pooling layer, a splicing layer, a 1×1 convolutional block, a 3×3 convolutional block, and a sigmoid activation function connected in sequence. The multi-scale channel enhancement branch includes a parallel average pooling layer, a 1×1 convolutional block, a 3×3 convolutional block, and a 1×1 convolutional block connected in sequence.

[0022] The outputs of the spatial attention branch and the multi-scale channel enhancement branch are weighted element-wise, and then added to the input features of the multi-branch multi-scale fusion attention mechanism module through residual connections to obtain the output of the multi-branch multi-scale fusion attention mechanism module.

[0023] Furthermore, step 4 also includes: performing preliminary screening on the detection boxes output by the remote sensing target detection model based on a preset confidence threshold to obtain candidate detection boxes; performing non-maximum suppression operation on the candidate detection boxes, calculating the degree of overlap between candidate detection boxes, suppressing redundant overlapping boxes within the same target area, and retaining the bounding box with the highest confidence as the target bounding box.

[0024] Secondly, the present invention provides a lightweight remote sensing image target detection system suitable for edge devices, comprising:

[0025] The dataset acquisition unit is used to acquire remote sensing image datasets and preprocess them to obtain training and test sets.

[0026] The model building unit is used to construct a remote sensing target detection model. This model uses the YOLOv8 network as a baseline model and incorporates a feature selection aggregation downsampling module, a heterogeneous receptive field feature enhancement module, and a multi-branch multi-scale fusion attention mechanism module. The aggregation downsampling module enhances the expressive power of shallow features; the heterogeneous receptive field feature enhancement module improves the robustness and directional consistency of feature representation; and the multi-branch multi-scale fusion attention mechanism module enhances the sensitivity and robustness of the remote sensing target detection model to targets in complex environments.

[0027] The model training unit is used to perform end-to-end training of the remote sensing target detection model using the training set.

[0028] The target detection unit is used to input the test set into the trained remote sensing target detection model and output the target bounding box and category prediction results.

[0029] Thirdly, the present invention provides an electronic device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the program, implements the method as described in the first aspect.

[0030] Fourthly, the present invention provides a non-transitory computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the method described in the first aspect.

[0031] The beneficial effects of this invention are as follows:

[0032] This invention proposes a lightweight remote sensing image target detection method suitable for edge devices, significantly improving detection accuracy and efficiency. The remote sensing image target detection model improves upon YOLOv8 as a baseline model. First, a feature selection aggregation downsampling module is introduced into the backbone network, effectively replacing traditional strided convolution and pooling, expanding the receptive field while reducing detail loss. Second, a designed heterogeneous receptive field feature enhancement structure fuses heterogeneous receptive field features, enhancing the modeling ability for multi-scale targets, especially small and dense targets, and improving the robustness of feature representation. A multi-branch, multi-scale fusion attention mechanism module is introduced into the neck network, adaptively focusing on high-response regions, strengthening the representation of key targets, suppressing background noise, and reducing the risk of missed and false detections. The overall network structure is compact, with few parameters, low computational complexity, and excellent inference speed, making it suitable for resource-constrained edge computing and UAV platforms. Through module collaborative optimization, this method significantly reduces model redundancy and resource consumption while maintaining high accuracy, effectively resolving the contradiction between accuracy and efficiency in traditional remote sensing detection, and possessing good engineering application value and promising prospects for widespread application. Attached Figure Description

[0033] Figure 1 A flowchart illustrating a lightweight remote sensing image target detection method suitable for edge devices provided in an embodiment of the present invention;

[0034] Figure 2 This is a structural diagram of the remote sensing target detection model provided in an embodiment of the present invention;

[0035] Figure 3 This is a design diagram of the feature aggregation downsampling module provided in an embodiment of the present invention;

[0036] Figure 4 This is a design diagram of a heterogeneous receptive field feature enhancement structure module provided in an embodiment of the present invention;

[0037] Figure 5 This is a design diagram of a multi-branch, multi-scale fusion attention mechanism module provided in an embodiment of the present invention;

[0038] Figure 6 This is a structural diagram of a lightweight remote sensing image target detection system suitable for edge devices, provided by an embodiment of the present invention.

[0039] Figure 7 This is a structural diagram of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0040] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of the embodiments of this invention will be clearly described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0041] like Figure 1 As shown, this embodiment of the invention provides a lightweight remote sensing image target detection method suitable for edge devices, including:

[0042] S1. Obtain the remote sensing image dataset and preprocess it to obtain the training set and test set.

[0043] Specifically, raw images captured by remote sensing satellites or drones are collected, and a remote sensing image dataset is constructed. Targets are labeled in the raw images, and the dataset is divided into training and testing sets in an 8:2 ratio. To standardize the input format, the images are normalized in size, and enhancement strategies such as random cropping, color perturbation, and mirror flipping are employed to improve the model's generalization ability.

[0044] S2. Construct a remote sensing target detection model. This model uses the YOLOv8 network as the baseline and introduces a feature selection aggregation downsampling module, a heterogeneous receptive field feature enhancement module, and a multi-branch multi-scale fusion attention mechanism module. The aggregation downsampling module enhances the expressive power of shallow features; the heterogeneous receptive field feature enhancement module improves the robustness and directional consistency of feature representation; and the multi-branch multi-scale fusion attention mechanism module enhances the sensitivity and robustness of the remote sensing target detection model to targets in complex environments.

[0045] S3. Use the training set to perform end-to-end training on the remote sensing target detection model.

[0046] Specifically, the remote sensing image training set constructed in step S1 and its corresponding target bounding box data are converted into vector form and input into the lightweight remote sensing image target detection model suitable for edge devices constructed in step S2 for end-to-end training, and the trained network model parameters are saved.

[0047] S4. Input the test set into the trained remote sensing target detection model and output the target bounding box and category prediction results.

[0048] Specifically, the original remote sensing images in the test set obtained in step S1 are input into the remote sensing image target detection method network model for edge devices that has been trained in step S3 to obtain accurate target detection results for subsequent evaluation.

[0049] The method provided in this invention enhances multi-scale feature extraction capabilities and adaptively improves key region representation by introducing a feature selection aggregation downsampling module and a heterogeneous receptive field feature enhancement module into the backbone network, and adding a multi-branch, multi-scale fusion attention mechanism module into the neck network. This method effectively addresses issues such as complex backgrounds, dense small targets, and significant scale variations in remote sensing images. While improving detection accuracy, it significantly reduces the number of model parameters and computational complexity, making it suitable for resource-constrained scenarios such as UAV platforms, satellite terminals, and edge computing devices.

[0050] As one possible implementation method, the remote sensing target detection model in step S2 has the following specific structure:

[0051] like Figure 2 As shown, the remote sensing target detection model uses the YOLOv8 network as the baseline model and introduces a feature selection aggregation downsampling module, a heterogeneous receptive field feature enhancement module, and a multi-branch multi-scale fusion attention mechanism module. Specifically, in the backbone network of the YOLOv8 network, the CBS module and C2F module in the feature extraction layer structure are replaced with the feature selection aggregation downsampling FSAD module and the heterogeneous receptive field feature enhancement C2F-HRFEM module, respectively. A multi-branch multi-scale fusion attention mechanism MMAM module is added at the connection between the neck network and the backbone network.

[0052] like Figure 3 As shown, the Feature Selection Aggregation Downsampling (FSAD) module is used to improve the expressive power of shallow features. The specific processing flow is as follows:

[0053] The input features of the feature selection aggregation downsampling module are processed by a 3×3 convolution to obtain the first feature;

[0054] The first feature is input into the SPDConv module for processing to obtain the second feature;

[0055] The first feature is processed by the dilated convolutional block and the SPDConv module to obtain the third feature;

[0056] The second and third features are concatenated by the concatenation unit and then input into the average pooling block and the max pooling block, respectively, to obtain the fourth and fifth features.

[0057] The fourth and fifth features are processed through convolutional blocks and the sigmoid function to obtain attention weights;

[0058] The second and third features are weighted by channel multiplication using attention weights. The second and third features after channel multiplication weighting are then fused and processed by 1×1 convolution to obtain the output features of the feature selection aggregation downsampling module.

[0059] Specifically, to address the problem of lost fine-grained feature information in remote sensing images, a feature selection aggregation downsampling module is designed to enhance the expressive power of shallow features. This involves converting the remote sensing image into vector form. First, local detail features are extracted using a standard 3×3 convolution, followed by dilated convolution to capture contextual information. Then, the output features of both the standard and dilated convolutions are transformed using a spatial-to-channel transformation mechanism (SPDConv), dividing the spatial dimension into multiple sub-images and concatenating them along the channel dimension to achieve information compression and enhancement. Finally, the two SPDConv-transformed features are concatenated and fused, effectively strengthening the model's ability to model edge textures and spatial structures. The formula is:

[0060]

[0061]

[0062]

[0063] in, This represents the input features of the feature selection aggregation downsampling module. This represents a 3×3 convolution. This indicates the SPDConv module, which uses a space-to-channel conversion mechanism; Indicates the second characteristic, This represents a 3×3 dilated convolution. Indicates the third characteristic, This represents the feature resulting from the concatenation of the second and third features. This indicates a splicing operation.

[0064] Subsequently, Global average pooling (GAP) and global max pooling (GMP) operations are performed to extract the fused global statistical features and local salient features. The results of GAP and GMP are concatenated and input into a 1×1 convolutional layer, and finally a channel attention weight vector is generated by the sigmoid activation function. Its size and and The same applies. Finally, the attention weight vector is applied to both branches, and the two weighted feature maps are added together to obtain the final downsampled output feature, as shown in the formula:

[0065]

[0066]

[0067]

[0068]

[0069] in, The symbol represents the sigmoid activation function, and ⊙ represents the channel-wise dot product operation. This represents the output features of the feature selection aggregation downsampling module.

[0070] like Figure 4 As shown, the heterogeneous receptive field feature enhancement C2F-HRFEM module is used to enhance the robustness and directional consistency of feature representation. Specifically, it includes: 3×3 convolutional blocks, parallel directional convolutional branches, and dynamic weight adjustment branches. The directional convolutional branches include a first branch and a second branch. The first branch includes a horizontal convolutional block, a vertical convolutional block, and a 1×1 convolutional block connected in sequence, and the second branch includes a vertical convolutional block, a horizontal convolutional block, and a 1×1 convolutional block connected in sequence. The dynamic weight adjustment branch includes a 3×3 convolutional block, an activation function layer, a 1×1 convolutional block, another activation function layer, a 1×1 convolutional block, a binary classification function, and a weight value layer connected in sequence. The outputs of the directional convolutional branches and the dynamic weight adjustment branches are weighted and fused, and then added to the input features of the heterogeneous receptive field feature enhancement module using a residual method to obtain the output of the heterogeneous receptive field feature enhancement module.

[0071] Specifically, targets in remote sensing images exhibit significant differences in size and orientation. For example, buildings, vehicles, and roads have structural features that differ significantly in texture directionality and scale. To improve the model's ability to model targets with different orientations and at multiple scales, a heterogeneous receptive field enhancement module with separable convolution and dynamically adjusted weights is introduced in the feature extraction stage to improve the robustness and directional consistency of feature representation. Figure 4 As shown, heterogeneous receptive field feature enhancement first involves input feature maps... The basic representation extracted by the input to a standard 3×3 convolution is denoted as 𝐹0. Subsequently, the features are fed in parallel into two directional convolution branches to capture structural information in different directions. The first branch (Branch-HV) employs an asymmetric directional convolution structure consisting of horizontal (1×3) and vertical (3×1) convolutions, aiming to enhance the expression of texture and edges in the horizontal direction; a 1×1 convolution is introduced after the directional convolutions to reshape features between channels. The second branch (Branch-VH) first uses a vertical convolution (3×1), then a horizontal convolution (1×3) to form an asymmetric convolution structure in the reverse direction, supplementing the vertical structure perception capability and enhancing the model's sensitivity to vertical texture changes; similarly, a 1×1 convolution is added at the end for channel adjustment. The features extracted by the two directional branches are concatenated along the channel dimension to form a set of directionally enhanced features. Simultaneously, the input features... The data is fed into a dynamic weighted subnet, which consists of consecutive 3×3 convolutions, activation functions, 1×1 convolutions, activation functions, and 1×1 convolutions. The output is an adaptive weight used to measure the importance of the two directional branches, so as to achieve fine control of the structure in different directions.

[0072] Finally, the directional branch features are weighted and fused under the guidance of dynamic weights, and then added to the initial input features through residual connections to obtain the final output of the heterogeneous receptive field feature enhancement module, thus achieving direction-sensitive and multi-scale consistent feature enhancement. The process can be represented as follows:

[0073]

[0074]

[0075]

[0076]

[0077]

[0078]

[0079] in, This represents the characteristics of the first branch output. This represents a 1×3 horizontal convolution. This represents a 3×1 vertical convolution. Represents a 1×1 convolution. This indicates the characteristics of the second branch output. This represents the activation function. This indicates that the output weights of the branches are dynamically adjusted. This represents the output feature of the heterogeneous receptive field feature enhancement module.

[0080] like Figure 5 As shown, the Multi-Branch Multi-Scale Fusion Attention Mechanism (MMAM) module is used to improve the sensitivity and robustness of the target detection model to targets in complex environments. Specifically, it includes a spatial attention branch and a multi-scale channel enhancement branch. The spatial attention branch contains a series of parallel average pooling layers and max pooling layers, a splicing layer, a 1×1 convolutional block, a 3×3 convolutional block, and a sigmoid activation function. The multi-scale channel enhancement branch contains a series of parallel average pooling layers, a 1×1 convolutional block, a 3×3 convolutional block, and a 1×1 convolutional block. The outputs of the spatial attention branch and the multi-scale channel enhancement branch are weighted element-wise and then added to the input features of the multi-branch multi-scale fusion attention mechanism module through residual connections to obtain the output of the multi-branch multi-scale fusion attention mechanism module.

[0081] Specifically, the spatial distribution and background features of targets in remote sensing images are often highly complex and uneven, especially in areas with varied terrain, severe occlusion, or low resolution, where target features exhibit significant local degradation and interference. To improve the sensitivity and robustness of target detection models in these complex environments, this invention introduces a multi-branch, multi-scale fusion attention mechanism module to effectively guide the model to focus on key regions and suppress background noise. For example... Figure 4 As shown, this module first inputs the feature map. The input is fed into two parallel branches: average pooling and max pooling, used to extract a global description from both statistical features and salient responses. The average pooling path obtains the overall distribution trend of features, while the max pooling path captures high-response regions to supplement local salient structural information. The two pooling results are concatenated after scale consistency and then channel compression is performed using a 1×1 convolution to reduce redundant feature dimensions. Subsequently, an activation function is applied to obtain a non-linear enhanced representation, and then convolution is used to generate a spatial attention map, which is normalized using the Sigmoid function to form soft weights for spatial dimension adjustment. Simultaneously, the input features... The data is then fed into a multi-scale channel enhancement branch. This branch first performs global average pooling on the features to extract channel descriptions that reflect the overall statistical significance. Subsequently, it sequentially passes through a multi-scale combination structure of 1×1 and 3×3 convolutions to achieve channel compression, local receptive field expansion, and structural pattern extraction. Finally, a 1×1 convolution restores the channel dimension, generating channel-enhanced features that fuse local and global patterns. This branch effectively captures multi-scale channel dependencies, thereby improving the expressive power of channel features.

[0082] Finally, the attention weights generated by the spatial attention branch are used to perform element-wise weighting of the multi-scale channel enhancement features, forming jointly attention-enhanced features. These enhanced features are then added to the input features via residual connections to obtain the final output. The process can be represented as follows:

[0083]

[0084]

[0085]

[0086]

[0087]

[0088]

[0089]

[0090]

[0091] in, Indicates average pooling. This indicates max pooling. This indicates element-wise multiplication. This represents the output feature after average pooling. This represents the output features after max pooling. The output of the spatial attention branch is indicated. This indicates the output of the multi-scale channel enhancement branch. This represents the output of the multi-branch, multi-scale fusion attention mechanism module.

[0092] As one possible implementation method, the specific process of step S3 is as follows:

[0093] S3.1. Convert the remote sensing image training set and its corresponding target bounding box data constructed in step S1 into vector form, and input them into the lightweight remote sensing image target detection method model suitable for edge devices constructed in step S2 for end-to-end training.

[0094] S3.2 Use PyCharm to write training scripts and train the model within the PyTorch deep learning framework. All experiments are implemented using PyTorch toolboxes and Python scripts. The network training uses the SGD optimizer, with an initial learning rate set to 1×10⁻⁶. -2 The weight decays to The momentum was set to 0.937, and a learning rate scheduling strategy was introduced to improve convergence stability. Training was conducted in an experimental environment equipped with an RTX 4090 GPU, with a total of 200 training epochs to ensure the model fully learned image features.

[0095] S3.3. Iterate to the specified number of rounds, stop updating, and save the trained network model parameters.

[0096] As one possible implementation, S4 further includes: performing preliminary screening on the detection boxes output by the remote sensing target detection model based on a preset confidence threshold to obtain candidate detection boxes; performing non-maximum suppression on the candidate detection boxes, calculating the degree of overlap between candidate detection boxes, and suppressing redundant overlapping boxes within the same target region, retaining the bounding box with the highest confidence as the target bounding box. Therefore, the specific process of S4 is as follows:

[0097] S4.1. The original remote sensing images in the test set constructed in step S1 are uniformly resized to conform to the input specifications of the training model in step S3. The resized images are normalized and tensorized, and then converted into a vector form readable by the network, which is then fed into the lightweight remote sensing image target detection model trained in step S3.

[0098] S4.2 Forward Inference and Feature Representation Generation. After receiving the input image, the model sequentially extracts and enhances shallow detail features and deep semantic features through the feature selection, aggregation, and downsampling module and the heterogeneous receptive field feature enhancement module in the backbone network. Then, it enters the neck network, where a multi-branch, multi-scale fusion attention mechanism module extracts channel importance weights, adaptively enhancing the feature representation of key target regions and suppressing background interference. Finally, the detection head generates the position offset, class probability, and confidence score of candidate target boxes.

[0099] S4.3 Candidate Box Selection and Threshold Filtering. The candidate boxes output by the model are first initially selected based on a preset confidence threshold, eliminating low-confidence prediction boxes to reduce redundant computation in subsequent processing and improve the quality of candidate target boxes. Non-maximum suppression (NMS) is then performed on the retained candidate boxes to calculate the overlap (IoU) between candidate boxes and suppress redundant overlapping boxes within the same object region, retaining only the bounding box with the highest confidence as the final detection box. The introduction of NMS effectively avoids duplicate detection, improving the consistency of target localization and display quality.

[0100] S4.3 The detection boxes filtered by NMS will be used as the final output results, including: target bounding box coordinates, target category label, confidence score. The above detection results can be used for visualization, accuracy assessment or further deployment in actual remote sensing business processes, such as land cover recognition, change detection, scene understanding and other tasks.

[0101] As one possible implementation, to verify the applicability and robustness of the lightweight remote sensing image target detection method for edge devices described in this invention under different remote sensing scenarios, this embodiment conducted comparative experiments with several mainstream target detection algorithms, including the YOLO series models (YOLOv3, YOLOv5L, YOLOX-L, YOLOv6, YOLOv8n) and a two-stage detection algorithm (Faster R-CNN), to comprehensively verify the effectiveness and advantages of this invention. As shown in Table 1, the lightweight remote sensing image target detection method for edge devices described in this invention achieved the highest mAP on the SIMD dataset, while consuming low computational resources, fully verifying its efficient detection capability and computational economy in complex remote sensing scenarios. This method has greater practical application value in environments with limited computational resources, and is especially suitable for edge computing platforms and embedded remote sensing systems.

[0102] Validated on typical SIMD remote sensing datasets, the method described in this invention demonstrates excellent performance in terms of detection accuracy, inference speed, and model lightweighting, proving its effectiveness and practical value in remote sensing target detection tasks. This invention not only enhances the automated processing capabilities in intelligent remote sensing image analysis but also provides feasible technical support for practical applications such as multi-source remote sensing information processing, ground feature monitoring, and emergency command.

[0103] Table 1 Performance comparison between the present invention and existing methods

[0104]

[0105] like Figure 6 As shown, this embodiment of the invention also provides a lightweight remote sensing target detection system suitable for edge devices, comprising:

[0106] The dataset acquisition unit is used to acquire remote sensing image datasets and preprocess them to obtain training and test sets.

[0107] The model building unit is used to construct a remote sensing target detection model. This model uses the YOLOv8 network as a baseline model and incorporates a feature selection aggregation downsampling module, a heterogeneous receptive field feature enhancement module, and a multi-branch multi-scale fusion attention mechanism module. The aggregation downsampling module enhances the expressive power of shallow features; the heterogeneous receptive field feature enhancement module improves the robustness and directional consistency of feature representation; and the multi-branch multi-scale fusion attention mechanism module enhances the sensitivity and robustness of the remote sensing target detection model to targets in complex environments.

[0108] The model training unit is used to perform end-to-end training of the remote sensing target detection model using the training set.

[0109] The target detection unit is used to input the test set into the trained remote sensing target detection model and output the target bounding box and category prediction results.

[0110] Figure 7 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 7 As shown, the electronic device may include a processor 701, a communications interface 702, a memory 703, and a communication bus 704. The processor 701, communications interface 702, and memory 703 communicate with each other via the communication bus 704. The processor 701 can call logical instructions stored in the memory 703 to execute a remote sensing image target detection method, which includes steps S1 to S4 of the above embodiments.

[0111] Furthermore, when the logical instructions in the aforementioned memory 703 are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0112] This invention also provides a computer program product, which includes a computer program stored on a non-transitory computer-readable storage medium. The computer program includes program instructions, and when the program instructions are executed by a computer, the computer can execute the remote sensing image target detection method provided in the above-described method embodiments, including steps S1 to S4 of the above embodiments.

[0113] This invention also provides a non-transitory computer-readable storage medium storing a computer program thereon. When the computer program is executed by a processor, it implements the remote sensing image target detection method provided in the above-described method embodiments, including steps S1 to S4 of the above embodiments.

[0114] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A lightweight remote sensing image target detection method suitable for edge devices, characterized in that, include: Step 1: Obtain the remote sensing image dataset and preprocess it to obtain the training set and test set; Step 2: Construct a remote sensing target detection model; The remote sensing target detection model described above uses the YOLOv8 network as the baseline model and introduces a feature selection aggregation downsampling module, a heterogeneous receptive field feature enhancement module, and a multi-branch multi-scale fusion attention mechanism module. The aggregation downsampling module is used to improve the expressive power of shallow features; the heterogeneous receptive field feature enhancement module is used to enhance the robustness and directional consistency of feature expression; the multi-branch multi-scale fusion attention mechanism module is used to improve the sensitivity and robustness of the remote sensing target detection model to targets in complex environments. Step 3: Perform end-to-end training of the remote sensing target detection model using the training set; Step 4: Input the test set into the trained remote sensing target detection model and output the target bounding box and category prediction results.

2. The lightweight remote sensing image target detection method suitable for edge devices according to claim 1, characterized in that, In step 1, the preprocessing includes uniformly normalizing the size of the remote sensing images in the remote sensing image dataset, then using a variety of online data augmentation strategies to augment the data, and dividing the training set and test set according to a preset ratio. The various online data augmentation strategies include: Mosaic mosaicking, color perturbation, random cropping, and mirroring.

3. The lightweight remote sensing image target detection method suitable for edge devices according to claim 1, characterized in that, In step 2, the remote sensing target detection model uses the YOLOv8 network as the baseline model and introduces a feature selection aggregation downsampling module, a heterogeneous receptive field feature enhancement module, and a multi-branch multi-scale fusion attention mechanism module. Specifically, it includes replacing the CBS module and C2F module in the backbone network of the YOLOv8 network with the feature selection aggregation downsampling module and the heterogeneous receptive field feature enhancement module, respectively, and adding the multi-branch multi-scale fusion attention mechanism module at the connection between the neck network and the backbone network.

4. The lightweight remote sensing image target detection method suitable for edge devices according to claim 1, characterized in that, In step 2, the feature selection aggregation downsampling module is used to improve the expressive power of shallow features. The specific processing flow is as follows: The input features of the feature selection aggregation downsampling module are processed by a 3×3 convolution to obtain the first feature; The first feature is input into the SPDConv module for processing to obtain the second feature; The first feature is processed by the dilated convolutional block and the SPDConv module to obtain the third feature; The second feature and the third feature are spliced ​​together by the splicing unit and then input into the average pooling block and the max pooling block respectively to obtain the fourth feature and the fifth feature; The fourth and fifth features are processed by convolutional blocks and a sigmoid function to obtain attention weights; The second feature and the third feature are weighted by channel multiplication using the attention weights respectively. The second feature and the third feature after channel multiplication weighting are then fused and processed by 1×1 convolution to obtain the output feature of the feature selection aggregation downsampling module.

5. A lightweight remote sensing image target detection method suitable for edge devices according to claim 1, characterized in that, In step 2, the heterogeneous receptive field feature enhancement module is used to enhance the robustness and directional consistency of feature representation, specifically including: a 3×3 convolutional block, parallel directional convolutional branches, and a dynamic weight adjustment branch; wherein the directional convolutional branch includes a first branch and a second branch, the first branch includes a horizontal convolutional block, a vertical convolutional block, and a 1×1 convolutional block connected in sequence, and the second branch includes a vertical convolutional block, a horizontal convolutional block, and a 1×1 convolutional block connected in sequence; the dynamic weight adjustment branch includes a 3×3 convolutional block, an activation function layer, a 1×1 convolutional block, an activation function layer, a 1×1 convolutional block, a binary classification function, and a weight value layer connected in sequence; The output of the heterogeneous receptive field feature enhancement module is obtained by weighted and fused with the outputs of the directional convolution branch and the dynamic weight adjustment branch, and then added to the input features of the heterogeneous receptive field feature enhancement module using a residual method.

6. The lightweight remote sensing image target detection method suitable for edge devices according to claim 1, characterized in that, In step 2, the multi-branch, multi-scale fusion attention mechanism module is used to improve the sensitivity and robustness of the target detection model to targets in complex environments. Specifically, it includes a spatial attention branch and a multi-scale channel enhancement branch. The spatial attention branch contains a series of parallel average pooling layers and max pooling layers, a splicing layer, a 1×1 convolutional block, a 3×3 convolutional block, and a sigmoid activation function. The multi-scale channel enhancement branch contains a series of parallel average pooling layers, a 1×1 convolutional block, a 3×3 convolutional block, and a 1×1 convolutional block. The outputs of the spatial attention branch and the multi-scale channel enhancement branch are weighted element-wise, and then added to the input features of the multi-branch multi-scale fusion attention mechanism module through residual connections to obtain the output of the multi-branch multi-scale fusion attention mechanism module.

7. The lightweight remote sensing image target detection method suitable for edge devices according to claim 1, characterized in that, Step 4 further includes: performing preliminary screening on the detection boxes output by the remote sensing target detection model based on a preset confidence threshold to obtain candidate detection boxes; performing non-maximum suppression operation on the candidate detection boxes, calculating the degree of overlap between candidate detection boxes, suppressing redundant overlapping boxes within the same target area, and retaining the bounding box with the highest confidence as the target bounding box.

8. A lightweight remote sensing target detection system suitable for edge devices, characterized in that, include: The dataset acquisition unit is used to acquire remote sensing image datasets and preprocess them to obtain training and test sets. Model building unit, used to build remote sensing target detection models; The remote sensing target detection model described above uses the YOLOv8 network as the baseline model and introduces a feature selection aggregation downsampling module, a heterogeneous receptive field feature enhancement module, and a multi-branch multi-scale fusion attention mechanism module. The aggregation downsampling module is used to improve the expressive power of shallow features; the heterogeneous receptive field feature enhancement module is used to enhance the robustness and directional consistency of feature expression; the multi-branch multi-scale fusion attention mechanism module is used to improve the sensitivity and robustness of the remote sensing target detection model to targets in complex environments. The model training unit is used to perform end-to-end training of the remote sensing target detection model using the training set. The target detection unit is used to input the test set into the trained remote sensing target detection model and output the target bounding box and category prediction results.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the method as described in any one of claims 1 to 7.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1 to 7.

Images