Remote sensing target detection method and device based on multi-scale and efficient attention mechanism

Abstract

The invention relates to a remote sensing target detection method and device based on multiple scales and an efficient attention mechanism, and the method comprises the steps: obtaining a target remote sensing image, inputting the target remote sensing image into a target detection model, and obtaining a target detection result; wherein the target detection model comprises the following steps: replacing a 3 * 3 common convolution in the sixth network layer ELANBlock of the YOLOv7 by using a 3 * 3 depth separable convolution; replacing a 3 * 3 common convolution in the ELANBlock of the eighth network layer of the YOLOv7 by using a 5 * 5 depth separable convolution; using a 7 * 7 depth separable convolution to replace a 3 * 3 common convolution in the ELANBlock of the tenth network layer of the YOLOv7; an MIA-M module based on multi-scale efficient attention is used for replacing the 17th network layer ELANBlock and the 23rd network layer ELANBlock of the YOLOv7; an MIA-S module based on multi-scale efficient attention is used for replacing the ELANBlock of the 21st network layer of the YOLOv7; and replacing the ELANBlock of the 25th network layer of the YOLOv7 with a multi-scale efficient attention MIA-L module to obtain a target detection model.

Description

Technical Field

[0001] The present invention belongs to the technical field of target detection, and in particular relates to a remote sensing target detection method and device based on a multi-scale and efficient attention mechanism. Background Art

[0002] As a non-contact, long-distance detection method, remote sensing technology has become increasingly valuable in the fields of aerial reconnaissance, environmental monitoring, military reconnaissance, resource exploration, etc. As an important part of remote sensing technology, remote sensing target detection has the core task of accurately identifying and locating the target object of interest from remote sensing images. However, remote sensing image target detection faces many challenges, which are mainly due to the particularity of remote sensing images.

[0003] First, the background of remote sensing images is complex and changeable. Since remote sensing images usually cover a wide geographical area, they contain a wide variety of elements such as natural landscapes, urban buildings, and transportation facilities. The boundary between the background and the target is often not clear, which increases the difficulty of target detection.

[0004] Secondly, the inter-class similarity in remote sensing images is high. Objects of different categories may have similar shape, texture, and color features, making it difficult for the detector to accurately distinguish them. At the same time, objects of the same category may also present completely different appearances under different lighting, angles, and resolutions, further increasing the difficulty of detection.

[0005] Furthermore, there is a multi-scale transformation problem for objects in remote sensing images. Due to changes in the shooting angle and distance of remote sensing images, the size and shape of the objects in the image will also change accordingly. This multi-scale transformation requires the detector to have the ability to accurately identify objects at different scales.

[0006] In addition, the targets in remote sensing images are usually small and dense. In high-resolution remote sensing images, multiple targets may be closely arranged together to form a dense target group. In this case, the detector needs to be able to accurately segment and identify each target to avoid missed detection and false detection.

[0007] In order to solve these problems, deep learning technology has been widely used in remote sensing image target detection. In recent years, with the continuous development of deep learning technology, many excellent target detectors have emerged, such as the YOLO series, Faster R-CNN, etc. These detectors have achieved remarkable results on natural images, but their performance on remote sensing images is not satisfactory. This is mainly because there are significant differences between remote sensing images and natural images in terms of feature distribution, background complexity, etc.

[0008] In order to improve the performance of remote sensing image target detection, researchers have begun to try to introduce new technologies such as attention mechanisms and Vision transformers into target detection models. For example, the introduction of attention mechanisms such as SEnet, ECA, and CBAM enables the model to pay more attention to important areas in the image and improve detection accuracy. At the same time, the emergence of Vision transformers also provides new ideas for target detection. It uses the self-attention mechanism to capture long-distance dependencies and global information in the image, further improving detection performance.

[0009] However, although these new technologies have improved the accuracy of remote sensing image target detection to a certain extent, some problems still exist. In particular, the computational complexity of the self-attention mechanism is high, which leads to a slow inference speed of the model and cannot meet the requirements of real-time detection. Therefore, how to improve the inference speed of the model while maintaining accuracy has become an urgent problem to be solved in the field of remote sensing image target detection. Summary of the invention

[0010] In order to solve the problems existing in the background technology, one aspect of the present invention provides a remote sensing target detection method based on a multi-scale and efficient attention mechanism, including: acquiring a target remote sensing image, inputting the target remote sensing image into a target detection model, and obtaining a target detection result; wherein the target detection model includes: replacing the 3×3 ordinary convolution in the 6th network layer ELANBlock of YOLOv7 with a 3×3 depth separable convolution; replacing the 3×3 ordinary convolution in the 8th network layer ELANBlock of YOLOv7 with a 5×5 depth separable convolution; using a 7×7 The 3×3 ordinary convolution in the 10th network layer ELANBlock of YOLOv7 is replaced by the depthwise separable convolution; the 17th network layer ELANBlock and the 23rd network layer ELANBlock of YOLOv7 are replaced by the MIA-M module based on multi-scale efficient attention; the 21st network layer ELANBlock of YOLOv7 is replaced by the MIA-S module based on multi-scale efficient attention; the 25th network layer ELANBlock of YOLOv7 is replaced by the MIA-L module based on multi-scale efficient attention to obtain the target detection model.

[0011] Another aspect of the present invention provides a remote sensing target detection method and device based on a multi-scale and efficient attention mechanism, comprising a processor and a memory; the memory is used to store a computer program; the processor is connected to the memory, and is used to execute the computer program stored in the memory, so that the remote sensing target detection device based on a multi-scale and efficient attention mechanism performs the remote sensing target detection method based on a multi-scale and efficient attention mechanism.

[0012] Another aspect of the present invention provides a computer-readable storage medium storing a program, characterized in that when the program is executed by a processor, it implements the remote sensing target detection method based on multi-scale and efficient attention mechanism.

[0013] The present invention has at least the following beneficial effects

[0014] The present invention strengthens the semantic and contour information in remote sensing image information and improves the detection accuracy by designing MIA-S module, MIA-M module and MIA-L module based on multi-scale efficient attention, and ELANBlock based on deep separable convolution. These modules can use different receptive fields for objects of different sizes, effectively process remote sensing images with severe scale effects, and solve the problem of high similarity between targets and backgrounds and susceptibility to interference from similar backgrounds. Compared with traditional self-attention modules (such as swin transformer, vit, etc.), the MIA module reduces the complexity of calculation, so that the model improves the detection accuracy under the premise of meeting real-time requirements. At the same time, replacing ordinary convolution with deep separable convolution also further improves the training speed and reasoning speed of the model. The present invention gives the remote sensing image detector the potential to cope with various complex scenes, improves the detection performance of the remote sensing image detector, and solves the problems of severe scale effects of remote sensing images and high similarity between targets and backgrounds. In summary, the present invention significantly improves the accuracy and adaptability of remote sensing image target detection while maintaining real-time performance, and provides a new and effective method for remote sensing image target detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 It is a model structure diagram of the present invention;

[0016] Figure 2 It is a schematic diagram of the structure of ELANBlock of the present invention;

[0017] Figure 3 It is a schematic diagram of the structure of the MIA-S module, the MIA-M module and the MIA-L module of the present invention;

[0018] Figure 4 It is a structural schematic diagram of the IAEF module of the present invention. DETAILED DESCRIPTION

[0019] The following describes the embodiments of the present invention by specific examples, and those skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments only illustrate the basic concept of the present invention in a schematic manner, and the following embodiments and features in the embodiments can be combined with each other without conflict.

[0020] See also Figure 1 to Figure 4 The present invention provides a remote sensing target detection method based on multi-scale and efficient attention mechanism, comprising:

[0021] Acquire a target remote sensing image, input the target remote sensing image into a target detection model, and obtain a target detection result; wherein the target detection model includes: replacing the 3×3 ordinary convolution in the 6th network layer ELANBlock of YOLOv7 with a 3×3 depthwise separable convolution; replacing the 3×3 ordinary convolution in the 8th network layer ELANBlock of YOLOv7 with a 5×5 depthwise separable convolution; replacing the 3×3 ordinary convolution in the 10th network layer ELANBlock of YOLOv7 with a 7×7 depthwise separable convolution; replacing the 17th network layer ELANBlock and the 23rd network layer ELANBlock of YOLOv7 with a MIA-M module based on multi-scale efficient attention; replacing the 21st network layer ELANBlock of YOLOv7 with a MIA-S module based on multi-scale efficient attention; replacing the 25th network layer ELANBlock of YOLOv7 with a MIA-L module based on multi-scale efficient attention, and obtaining a target detection model.

[0022] The dataset used in the embodiment of the present invention is composed of the public general remote sensing target detection dataset DOTAv1.0, RSOD and NWPU-VHR10. The model framework is built by pytorch. The training configuration of the model is the same as the baseline tracker YOLO series. The batch size is set to 6, and 300 epochs are trained without pre-training parameters. During training, the image size is adjusted to 1024×1024×3.

[0023] In the backbone network, the downsampled feature maps are processed with 7×7, 5×5, and 3×3 convolution kernels in the improved ELAN network structure respectively; the downsampled image is processed by a multi-level fusion feature pool to obtain a feature map of 32×32×1024 size with detail enhancement, and the feature map is upsampled in the neck network. The feature maps of the same resolution in the backbone network are fused, and the proposed attention module is used for feature extraction. At the same time, for feature maps of different sizes, efficient hybrid attention modules with different receptive fields are used for feature extraction to obtain feature maps with better target texture.

[0024] Preferably, the 8th network layer ELANBlock and the 10th network layer ELANBlock of YOLOv7 have the same network structure, both including: three 1×1 ordinary convolutions and four 3×3 ordinary convolutions; wherein, the first 1×1 ordinary convolution and the second 1×1 ordinary convolution are respectively used to extract features from the input feature map to obtain features f1 and f2; the first 3×3 ordinary convolution is used to extract features from feature f2 to obtain feature x1, the second 3×3 ordinary convolution is used to extract features from feature x1 to obtain feature x2, the third 3×3 ordinary convolution is used to extract features from feature x2 to obtain feature x3; the fourth 3×3 ordinary convolution is used to extract features from feature x3 to obtain feature x4; features f1, f2, x2 and x4 are concatenated to obtain feature f3, and feature f3 is input into the third 1×1 ordinary convolution to obtain the output feature of ELANBlock.

[0025] Preferably, the replacing the 3×3 ordinary convolution in the 6th network layer ELANBlock of YOLOv7 with the 3×3 depthwise separable convolution comprises: replacing the first 3×3 ordinary convolution, the third 3×3 ordinary convolution and the fourth 3×3 ordinary convolution in the 6th network layer ELANBlock of YOLOv7 with 3×3 depthwise separable convolution.

[0026] See also Figure 2 Preferably, the replacing of the 3×3 ordinary convolution in the 8th network layer ELANBlock of YOLOv7 with the 5×5 depthwise separable convolution comprises: replacing the first 3×3 ordinary convolution, the third 3×3 ordinary convolution and the fourth 3×3 ordinary convolution in the 8th network layer ELANBlock of YOLOv7 with 5×5 depthwise separable convolution.

[0027] Preferably, replacing the 3×3 ordinary convolution in the 10th network layer ELANBlock of YOLOv7 with the 7×7 depthwise separable convolution includes: replacing the first 3×3 ordinary convolution, the third 3×3 ordinary convolution and the fourth 3×3 ordinary convolution in the 10th network layer ELANBlock of YOLOv7 with 7×7 depthwise separable convolution.

[0028] See also Figure 3 Preferably, the MIA-S module, the MIA-M module and the MIA-L module are multi-scale efficient attention modules with successively larger scales, each of which includes: two 1×1 conventional convolutions, three 3×3 conventional convolutions and an efficient attention IAEF module; wherein the two 1×1 conventional convolutions are respectively used to perform feature extraction on the input feature map to obtain features y1 and y2; the efficient attention IAEF module is used to perform an efficient attention operation on feature y2 to obtain feature y3; the first 3×3 conventional convolution is used to perform feature extraction on feature y3 to obtain feature y4; the second 3×3 conventional convolution is used to perform feature extraction on feature y4 to obtain y5; the third 3×3 conventional convolution is used to perform feature extraction on feature y5 to obtain y6; the features y1, y2, y3, y4, y5 and y6 are concatenated to obtain the output feature.

[0029] See also Figure 4 Preferably, the efficient attention IAEF module is used to perform an efficient attention operation on the feature y2 to obtain the feature y3, including:

[0030] y3=α×attn+β×Conv

[0031]

[0032] K′=Conv 3×3,groups＝chann (K)

[0033] Conv=DWConv k×k (Conv 1×1 (Cat(Q,K,V))

[0034] Q=Conv 1×1 (y2)

[0035] K=Conv 1×1 (y2)

[0036] V=Conv 1×1 (y2)

[0037] Among them, α and β are two learnable parameters; Conv 1×1 (Cat(Q,K′)).mean dim＝channe Indicates that Conv1×1 (Cat(Q,K′)) performs mean operation on the channel dimension; Conv 3×3,groups＝channe represents 3×3 grouped convolution; Cat() represents the concatenation function; DWConv k×k represents a k×k depthwise separable convolution.

[0038] In this embodiment, the input feature map of efficient attention and depth-separable convolution comes from three 1×1 convolutions. After the input feature map is convolved, three feature matrices are obtained, which are denoted as Q, K, and V.

[0039] First, the K matrix is ​​processed using a depth-separable convolution with a convolution kernel size of 3×3 to obtain the feature matrix K′ after information enhancement. Then, the Q and K′ matrices are spliced ​​along the channel dimension, and then a 1×1 convolution is performed to realize the information communication between the channel dimensions of the feature matrix. The result is cross-multiplied with the V matrix, and finally a residual link is made with K′, which effectively provides an additional information transmission path, so that the network can make more comprehensive use of the input information. While the efficient attention module is running, Q, K, and V are spliced ​​along the channel dimension and input into a depth-separable convolution. The size of the convolution kernel changes with the change of the depth in the neck network. For the feature map corresponding to the large target, a 7×7 convolution kernel is used; for the feature map corresponding to the medium target, a 5×5 convolution kernel is used; for the feature map corresponding to the large target, a 3×3 convolution kernel is used. Finally, a weighted fusion is performed on the feature maps extracted by the two modules. After extracting the features of large, medium, and small scales through the MIA module in the neck network, feature extraction is performed using reparameterized convolution. For image data, it can capture both local detail features (through small-scale convolution branches) and more macroscopic regional features (through large-scale convolution branches), making the model's feature expression of the data more comprehensive and accurate.

[0040] Preferably, in the DWConv of the MIA-S module k×k The k value in the MIA-M module is 3. k×k The k value in the MIA-L module is 5. k×k The value of k in is 7.

[0041] During the training process, the output features are processed according to the feature maps representing large, medium, and small targets for classification. The dimensions of the feature maps are batchSize×20×20×(cls+4+conf), batchSize×40×40×(cls+4+conf), batchSize×80×80×(cls+4+conf), where batchsize represents the number of samples input for one training, cls represents the number of types in the data set, and conf represents the confidence of the type output by the model. In training mode, CIoU and 01 binary cross entropy losses are calculated based on the predicted coordinates, types, and confidence information. The final loss function is the weighted sum of the three loss functions, expressed as:

[0042] Loss total =λ cls Loss cls +λ box Loss box +λ obj Loss obj

[0043] After experimental adjustment, λ cls ,λ box ,λ obj The values ​​of the three weights are set to 0.3, 0.05, and 0.7 respectively.

[0044] In inference mode, the dimensions of the feature maps will be adjusted. The dimensions of the three feature maps corresponding to large, medium and small targets will be adjusted to batchsize*3072*(cls+conf+4), batch*12288*(cls+conf+4) and batchsize*49152*(cls+conf+4). Then, according to the NMS non-maximum suppression algorithm, the predicted content is processed to avoid redundant target boxes in the predicted map, and finally the detection results are output.

[0045] During the training process, the model's location information is the center coordinates of the prediction box and the width and height of the prediction box.

[0046] It can be expressed as:

[0047] box pred =(x center ,y cente ,t w ,t h )

[0048] Among them, Xcenter, Ycenter are the center coordinates of the prediction box, Tw, Th are the width and height of the prediction box.

[0049] The prediction box restoration formula under feature maps of different resolutions conforms to the following formula:

[0050] b x =(2σ(t x )-0.5+c x )*stride

[0051] b y =(2σ(t y )-0.5+c y )*stride

[0052] b w =((2σ(t w )) 2 *anchor_grid

[0053] b h =((2σ(t h )) 2 *anchor_grid

[0054] where b x , b y , b w , b h They represent the center coordinates and width and height of the prediction box in the actual image respectively; σ represents the sigmoid activation function, so that the offset is normalized to [0,1]; σ represents the sigmoid activation function, Cx, Cy represent the x, y coordinates of the center of the prediction box, Tw, Th represent the width and height of the prediction box; stride represents the downsampling multiple; anchor_grid represents the product of the resolution of the feature map after the current downsampling and the stride.

[0055] In the loss function, the CIoU loss is calculated by the coordinate information of the predicted box and the actual target box, which can be expressed as:

[0056]

[0057] Among them, ρ 2 (β,β gt ) represents the square of the Euclidean distance between the center point of the predicted box and the real box, c represents the diagonal of the minimum enclosing rectangle of the predicted box and the real box; α is a weight coefficient used to balance the distance-related terms and the aspect ratio-related terms; IoU represents the intersection area of ​​the predicted box and the real box divided by their union area.

[0058] In the loss function, the classification loss is calculated using the cross entropy loss. For each prediction box, if the prediction box is assigned to a real target, the cross entropy loss between the probability distribution predicted by the prediction box and the real category can be expressed as:

[0059]

[0060] Among them, n represents the total number of categories, y represents the true category, and the predicted category probability distribution is p = (p1,

[0061] p2, p3...p n ).

[0062] In the loss function, the target confidence loss uses binary cross entropy loss. It is calculated as:

[0063] L obj =-y obj log(p obj )-(1-y obj )log(1-p ob )

[0064] The final loss function is the weighted sum of the three loss functions, expressed as:

[0065] Loss total =λ cls Loss cls +λ box Loss box +λ obj Loss obj

[0066] Among them, λ cls ,λ box ,λ obj They are the weight coefficients of classification loss, bounding box loss, and target confidence loss, which are used to adjust the relative importance of each loss part in the total loss.

[0067] Another aspect of the present invention provides a remote sensing target detection method and device based on a multi-scale and efficient attention mechanism, comprising a processor and a memory; the memory is used to store a computer program; the processor is connected to the memory, and is used to execute the computer program stored in the memory, so that the remote sensing target detection device based on a multi-scale and efficient attention mechanism performs the remote sensing target detection method based on a multi-scale and efficient attention mechanism.

[0068] Another aspect of the present invention provides a computer-readable storage medium storing a program, characterized in that when the program is executed by a processor, it implements the remote sensing target detection method based on multi-scale and efficient attention mechanism.

[0069] In summary, the present invention introduces MIA-S, MIA-M and MIA-L modules based on multi-scale efficient attention into the target detection model. The present invention can use different receptive fields for targets of different sizes and effectively deal with the scale effect problem in remote sensing images. Replacing the ordinary convolution in YOLOv7 with deep separable convolution not only reduces the amount of calculation, but also improves the generalization ability of the model, so that the model can more accurately identify targets under complex and changeable remote sensing image backgrounds. Although the attention mechanism is introduced, the present invention reduces the complexity of calculation by designing an efficient attention IAEF module, so that the model can meet the requirements of real-time detection while maintaining high precision. The use of deep separable convolution also further reduces the amount of calculation of the model and improves the reasoning speed. The introduction of the multi-scale efficient attention module enables the model to better deal with the problem of high similarity between the target and the background in the remote sensing image and susceptibility to interference from similar backgrounds, thereby improving the robustness of the model. By fusing features of different scales in the neck network, the model can capture richer information and further enhance the recognition ability of the model. This paper introduces multi-scale and efficient attention mechanisms, combined with deep separable convolution and re-parameterized convolution and other technical means, significantly improving the accuracy and real-time performance of remote sensing image target detection, and enhancing the robustness and adaptability of the model. This achievement provides new solutions and technical ideas for the field of remote sensing image target detection, which has important theoretical significance and practical application value.

[0070] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiment methods can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned methods. Among them, any reference to memory, storage, database or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration and not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), memory bus (Rambus) direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM).

[0071] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solution of the present invention can be modified or replaced by equivalents without departing from the purpose and scope of the technical solution, which should be included in the scope of the claims of the present invention.

Claims

1. A remote sensing target detection method based on multi-scale and efficient attention mechanism, characterized in that: include: Acquire a target remote sensing image, input the target remote sensing image into a target detection model, and obtain a target detection result; wherein the target detection model includes: replacing the 3×3 ordinary convolution in the 6th network layer ELANBlock of YOLOv7 with a 3×3 depthwise separable convolution; replacing the 3×3 ordinary convolution in the 8th network layer ELANBlock of YOLOv7 with a 5×5 depthwise separable convolution; replacing the 3×3 ordinary convolution in the 10th network layer ELANBlock of YOLOv7 with a 7×7 depthwise separable convolution; replacing the 17th network layer ELANBlock and the 23rd network layer ELANBlock of YOLOv7 with a MIA-M module based on multi-scale efficient attention; replacing the 21st network layer ELANBlock of YOLOv7 with a MIA-S module based on multi-scale efficient attention; replacing the 25th network layer ELANBlock of YOLOv7 with a MIA-L module based on multi-scale efficient attention, and obtaining a target detection model.

2. According to claim 1, a remote sensing target detection method based on multi-scale and efficient attention mechanism is characterized in that: The 8th network layer ELANBlock and the 10th network layer ELANBlock of YOLOv7 have the same network structure, both including: three 1×1 ordinary convolutions and four 3×3 ordinary convolutions; wherein, the first 1×1 ordinary convolution and the second 1×1 ordinary convolution are used to extract features from the input feature map to obtain features f1 and f2 respectively; the first 3×3 ordinary convolution is used to extract features from feature f2 to obtain feature x1, the second 3×3 ordinary convolution is used to extract features from feature x1 to obtain feature x2, the third 3×3 ordinary convolution is used to extract features from feature x2 to obtain feature x3; the fourth 3×3 ordinary convolution is used to extract features from feature x3 to obtain feature x4; features f1, f2, x2 and x4 are concatenated to obtain feature f3, and feature f3 is input into the third 1×1 ordinary convolution to obtain the output feature of ELANBlock.

3. The remote sensing target detection method based on multi-scale and efficient attention mechanism according to claim 2, characterized in that: The method of replacing the 3×3 ordinary convolution in the 6th network layer ELANBlock of YOLOv7 with the 3×3 depthwise separable convolution includes: replacing the first 3×3 ordinary convolution, the third 3×3 ordinary convolution and the fourth 3×3 ordinary convolution in the 6th network layer ELANBlock of YOLOv7 with 3×3 depthwise separable convolution.

4. The remote sensing target detection method based on multi-scale and efficient attention mechanism according to claim 2, characterized in that: The method of replacing the 3×3 ordinary convolution in the 8th network layer ELANBlock of YOLOv7 with the 5×5 depthwise separable convolution includes: replacing the first 3×3 ordinary convolution, the third 3×3 ordinary convolution and the fourth 3×3 ordinary convolution in the 8th network layer ELANBlock of YOLOv7 with 5×5 depthwise separable convolution.

5. The remote sensing target detection method based on multi-scale and efficient attention mechanism according to claim 2, characterized in that: The method of replacing the 3×3 ordinary convolution in the 10th network layer ELANBlock of YOLOv7 with the 7×7 depthwise separable convolution includes: replacing the first 3×3 ordinary convolution, the third 3×3 ordinary convolution and the fourth 3×3 ordinary convolution in the 10th network layer ELANBlock of YOLOv7 with 7×7 depthwise separable convolution.

6. The remote sensing target detection method based on multi-scale and efficient attention mechanism according to claim 1, characterized in that: The MIA-S module, MIA-M module and MIA-L module are multi-scale efficient attention modules with successively larger scales, each of which includes: two 1×1 conventional convolutions, three 3×3 conventional convolutions and an efficient attention IAEF module; wherein the two 1×1 conventional convolutions are respectively used to perform feature extraction on the input feature map to obtain features y1 and y2; the efficient attention IAEF module is used to perform an efficient attention operation on feature y2 to obtain feature y3; the first 3×3 conventional convolution is used to perform feature extraction on feature y3 to obtain feature y4; the second 3×3 conventional convolution is used to perform feature extraction on feature y4 to obtain y5; the third 3×3 conventional convolution is used to perform feature extraction on feature y5 to obtain y6; the features y1, y2, y3, y4, y5 and y6 are concatenated to obtain the output features.

7. The remote sensing target detection method based on multi-scale and efficient attention mechanism according to claim 6, characterized in that: The efficient attention IAEF module is used to perform efficient attention operation on feature y2 to obtain feature y3, including: y3=α×attn+β×Conv K′=Conv 3×3,groups＝chan (K) Conv=DWConv k×k (Conv 1×1 (Cat(Q,K,V)) Q=Conv 1×1 (y2) K=Conv 1×1 (y2) V=Conv 1×1 (y2) Among them, α and β are two learnable parameters; Conv 1×1 (Cat(Q,K′)).mean dim＝channel Indicates that Conv 1×1 (Cat(Q,K′)) performs mean operation on the channel dimension; Conv 3×3,groups＝channe represents 3×3 grouped convolution; Cat() represents the concatenation function; DWConv k×k represents a k×k depthwise separable convolution.

8. The remote sensing target detection method based on multi-scale and efficient attention mechanism according to claim 6, characterized in that: DWConv in MIA-S module k×k The k value in the MIA-M module is 3. k×k The k value in the MIA-L module is 5. k×k The value of k in is 7.

9. A remote sensing target detection device based on multi-scale and efficient attention mechanism, characterized in that: It includes a processor and a memory; the memory is used to store a computer program; the processor is connected to the memory and is used to execute the computer program stored in the memory, so that the remote sensing target detection device based on a multi-scale and efficient attention mechanism executes a remote sensing target detection method based on a multi-scale and efficient attention mechanism as described in any one of claims 1 to 7.

10. A computer-readable storage medium storing a program, characterized in that: When the program is executed by a processor, a remote sensing target detection method based on a multi-scale and efficient attention mechanism as described in any one of claims 1 to 7 is implemented.

Images