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

Abstract

This invention relates to a remote sensing target detection method and apparatus based on a multi-scale and efficient attention mechanism, comprising: acquiring a remote sensing image of the target; inputting the remote sensing image of the target into a target detection model; and obtaining a target detection result; wherein the target detection model comprises: replacing the ordinary 3×3 convolution in the 6th layer of the YOLOv7 network ELANBlock with a 3×3 depthwise separable convolution; replacing the ordinary 3×3 convolution in the 8th layer of the YOLOv7 network ELANBlock with a 5×5 depthwise separable convolution; and replacing the ordinary 3×3 convolution in the 8th layer of the YOLOv7 network ELANBlock with a 7×7 depthwise separable convolution. Replace the 3×3 ordinary convolution in the 10th layer ELANBlock of YOLOv7 with the MIA-M module based on multi-scale efficient attention; replace the 17th and 23rd layers ELANBlock of YOLOv7 with the MIA-S module based on multi-scale efficient attention; replace the 21st layer ELANBlock of YOLOv7 with the MIA-S module based on multi-scale efficient attention; replace the 25th layer ELANBlock of YOLOv7 with the MIA-L module based on multi-scale efficient attention to obtain the object detection model.

Description

Technical Field

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

[0002] Remote sensing technology, as a non-contact, long-range detection method, is increasingly valuable in fields such as aerial reconnaissance, environmental monitoring, military reconnaissance, and resource exploration. Remote sensing target detection, a crucial component of remote sensing technology, has the core task of accurately identifying and locating targets of interest from remote sensing images. However, target detection in remote sensing images faces numerous challenges, primarily stemming from the unique characteristics of remote sensing images.

[0003] First, the backgrounds of remote sensing images are complex and varied. Because remote sensing images typically cover vast geographical areas, 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 unclear, which increases the difficulty of target detection.

[0004] Secondly, remote sensing images exhibit high inter-class similarity. Targets of different categories may have similar shape, texture, and color features, making it difficult for detectors to accurately distinguish them. Furthermore, targets of the same category may appear drastically different under varying lighting, angles, and resolutions, further increasing the difficulty of detection.

[0005] Furthermore, targets in remote sensing images exhibit multi-scale transformation. Due to variations in the shooting angle and distance of remote sensing images, the size and shape of targets within the image also change accordingly. This multi-scale transformation necessitates that detectors possess the ability to accurately identify targets at different scales.

[0006] Furthermore, targets in remote sensing images are typically small and densely packed. In high-resolution remote sensing images, multiple targets may be closely arranged together, forming dense target clusters. In such cases, the detector needs to be able to accurately segment and identify each target, avoiding missed detections and false detections.

[0007] To address these challenges, deep learning technology has been widely applied to target detection in remote sensing images. In recent years, with the continuous development of deep learning technology, many excellent target detectors have emerged, such as the YOLO series and Faster R-CNN. While these detectors have achieved remarkable results on natural images, their performance on remote sensing images is less than satisfactory. This is mainly because remote sensing images differ significantly from natural images in terms of feature distribution and background complexity.

[0008] To improve the performance of object detection in remote sensing images, researchers have begun to introduce new technologies such as attention mechanisms and Vision Transformers into object detection models. For example, the introduction of attention mechanisms such as SENet, ECA, and CBAM allows the models to focus more on important regions in the image, improving detection accuracy. Meanwhile, the emergence of Vision Transformers has also provided a new approach to object detection, utilizing self-attention mechanisms to capture long-range dependencies and global information in the image, further enhancing detection performance.

[0009] However, while these new technologies have improved the accuracy of target detection in remote sensing images to some extent, some problems still exist. In particular, the computational complexity of the self-attention mechanism is high, resulting in slow inference speed and failing to 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] To address the problems existing in the background art, one aspect of the present invention provides a remote sensing target detection method based on multi-scale and efficient attention mechanisms, comprising: acquiring a remote sensing image of the target; inputting the remote sensing image of the target into a target detection model to obtain a target detection result; wherein, the target detection model comprises: replacing the ordinary 3×3 convolution in the 6th layer ELANBlock of YOLOv7 with a 3×3 depthwise separable convolution; replacing the ordinary 3×3 convolution in the 8th layer ELANBlock of YOLOv7 with a 5×5 depthwise separable convolution; and using a 7×7 depthwise separable convolution... The 3×3 ordinary convolution in the 10th layer of YOLOv7, ELANBlock, is replaced by a depthwise separable convolution; the 17th and 23rd layers of YOLOv7, ELANBlock, are replaced by the MIA-M module based on multi-scale efficient attention; the 21st layer of YOLOv7, ELANBlock, is replaced by the MIA-S module based on multi-scale efficient attention; and the 25th layer of YOLOv7, ELANBlock, is replaced by the MIA-L module based on multi-scale efficient attention, resulting in the object detection model.

[0011] Another aspect of the present invention provides a remote sensing target detection method apparatus based on 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 apparatus based on multi-scale and efficient attention mechanism performs the remote sensing target detection method based on 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 aforementioned remote sensing target detection method based on a multi-scale and efficient attention mechanism.

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

[0014] This invention enhances semantic and contour information in remote sensing images and improves detection accuracy by designing MIA-S, MIA-M, and MIA-L modules based on multi-scale efficient attention, and ELANBlock based on depthwise separable convolution. These modules can use different receptive fields for objects of different sizes, effectively handling remote sensing images with severe scale effects and solving the problems of high target-background similarity and susceptibility to interference from similar backgrounds. Compared to traditional self-attention modules (such as Swin Transformer and VIT), the MIA modules reduce computational complexity, allowing the model to improve detection accuracy while meeting real-time requirements. Simultaneously, replacing ordinary convolution with depthwise separable convolution further improves the model's training and inference speeds. This invention endows remote sensing image detectors with the potential to handle various complex scenarios, improves their detection performance, and solves problems such as severe scale effects and high target-background similarity in remote sensing images. In summary, this invention significantly improves the accuracy and adaptability of remote sensing image target detection while maintaining real-time performance, providing a new and effective method for target detection in remote sensing images. Attached Figure Description

[0015] Figure 1 This is a structural diagram of the model of the present invention;

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

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

[0018] Figure 4 This is a schematic diagram of the IAEF module of the present invention. Detailed Implementation

[0019] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can be modified or changed 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 are only schematic representations of the basic concept of the present invention. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0020] Please see Figures 1-4 This invention provides a remote sensing target detection method based on multi-scale and efficient attention mechanisms, comprising:

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

[0022] In this embodiment of the invention, the dataset used consists of the publicly available general remote sensing target detection datasets DOTAv1.0, RSOD, and NWPU-VHR10. A model framework was built using PyTorch, and the model's training configuration was the same as the baseline tracker YOLO series, with a batch size of 6 and training for 300 epochs, without pre-training parameters. During training, the image size was adjusted to 1024×1024×3.

[0023] In the backbone network, the downsampled feature maps are processed using 7×7, 5×5, and 3×3 convolutional kernels in the improved ELAN network structure, respectively. After multi-level fusion feature pooling, the downsampled images are processed to obtain feature maps of size 32×32×1024 with enhanced details. The feature maps are then upsampled in the neck network, and feature maps of the same resolution in the backbone network are fused. Feature extraction is performed using the proposed attention module. At the same time, efficient hybrid attention modules with different receptive fields are used to extract features from feature maps of different sizes, resulting in feature maps with more target texture.

[0024] Preferably, the 8th and 10th layers of the YOLOv7 network, ELANBlock, have the same network structure, both including: three 1×1 ordinary convolutions and four 3×3 ordinary convolutions; wherein, the first and second 1×1 ordinary convolutions are used to extract features f1 and f2 from the input feature map, respectively; the first 3×3 ordinary convolution is used to extract features x1 from feature f2, the second 3×3 ordinary convolution is used to extract features x2 from feature x1, the third 3×3 ordinary convolution is used to extract features x3 from feature x2, and the fourth 3×3 ordinary convolution is used to extract features x4 from feature x3; 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, replacing the 3×3 ordinary convolutions in the 6th layer ELANBlock of YOLOv7 with 3×3 depthwise separable convolutions includes replacing the first, third, and fourth ordinary convolutions in the 6th layer ELANBlock of YOLOv7 with 3×3 depthwise separable convolutions.

[0026] Please see Figure 2 Preferably, replacing the 3×3 ordinary convolutions in the 8th layer ELANBlock of YOLOv7 with 5×5 depthwise separable convolutions includes replacing the first, third, and fourth ordinary convolutions in the 8th layer ELANBlock of YOLOv7 with 5×5 depthwise separable convolutions.

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

[0028] Please see Figure 3 Preferably, the MIA-S module, MIA-M module, and MIA-L module are multi-scale efficient attention modules with progressively larger scales, each comprising: two 1×1 regular convolutions, three 3×3 regular convolutions, and an efficient attention IAEF module; wherein, the two 1×1 regular convolutions are used to extract features from the input feature map to obtain features y1 and y2 respectively; the efficient attention IAEF module is used to perform efficient attention operations on feature y2 to obtain feature y3; the first 3×3 regular convolution is used to extract features from feature y3 to obtain feature y4; the second 3×3 regular convolution is used to extract features from feature y4 to obtain feature y5; the third 3×3 regular convolution is used to extract features from feature y5 to obtain feature y6; and the features y1, y2, y3, y4, y5, and y6 are concatenated to obtain the output feature.

[0029] Please see Figure 4 Preferably, the efficient attention IAEF module is used to perform efficient attention operations on feature y2 to obtain 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] Here, α and β are two learnable parameters; Conv 1×1 (Cat(Q,K′)).mean dim＝channe Indicates Conv1×1 (Cat(Q,K′)) performs mean calculation along the channel dimension; Conv 3×3,groups＝channe This represents a 3×3 grouped convolution; Cat() represents the concatenation function; DWConv k×k This represents a k×k depth separable convolution.

[0038] In this embodiment, the input feature map for efficient attention and depth-separable convolution comes from three 1×1 convolutions. After convolution processing, the input feature map yields three feature matrices, denoted as Q, K, and V.

[0039] First, a depthwise separable convolution with a kernel size of 3×3 is used to process the K matrix, resulting in an information-enhanced feature matrix K′. Then, the Q and K′ matrices are concatenated along the channel dimension, followed by a 1×1 convolution to facilitate information exchange between the feature matrix channels. The result is then cross-productted with the V matrix, and finally, a residual link is established with K′, effectively providing additional information transmission paths and enabling the network to utilize input information more comprehensively. Simultaneously, while the efficient attention module is running, the concatenation of Q, K, and V along the channel dimension is input into a depthwise separable convolution. The kernel size varies with the depth of the neck network: a 7×7 kernel is used for feature maps corresponding to large targets; a 5×5 kernel is used for feature maps corresponding to medium targets; and a 3×3 kernel is used for feature maps corresponding to large targets. Finally, the feature maps extracted by the two modules are weighted and fused. In the neck network, features at large, medium, and small scales are extracted using the MIA module, and then reparameterized convolutions are used for feature extraction. For image data, it can capture both local detailed features (through small-scale convolution branches) and more macroscopic regional features (through large-scale convolution branches), making the model's feature representation of the data more comprehensive and accurate.

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

[0041] During training, the output features are classified according to feature maps representing large, medium, and small targets. The dimensions of the feature maps are batchSize×20×20×(cls+4+conf), batchSize×40×40×(cls+4+conf), and batchSize×80×80×(cls+4+conf), respectively, where batchSize represents the number of samples in one training iteration, cls represents the number of categories in the dataset, and conf represents the confidence score of the model's output category. In training mode, CIoU and 0 / 1 binary cross-entropy loss are calculated based on the predicted coordinates, categories, and confidence scores. The final loss function is a 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 adjustments, λ cls , λ box , λ obj The values ​​of the three weights were set to 0.3, 0.05, and 0.7, respectively.

[0044] In inference mode, the dimensions of the feature maps are adjusted. The dimensions of the feature maps for large, medium and small targets are adjusted to batchsize*3072*(cls+conf+4), batch*12288*(cls+conf+4) and batchsize*49152*(cls+conf+4), respectively. Then, the predicted content is processed by the Non-Maximum Suppression (NMS) algorithm to avoid redundant target boxes in the predicted map, and finally the detection result is output.

[0045] During training, the model's positional information consists of the center coordinates of the predicted bounding box and the width and height of the predicted bounding box.

[0046] It can be represented as:

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

[0048] Where Xcenter and Ycenter are the center coordinates of the predicted bounding box, and Tw and Th are the width and height of the predicted bounding box.

[0049] The formula for restoring the predicted bounding box 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 These represent the center coordinates and width and height of the predicted bounding box in the actual image, respectively; σ represents the sigmoid activation function, which normalizes the offset to [0,1]; σ represents the sigmoid activation function, Cx and Cy represent the x and y coordinates of the center of the predicted bounding box, Tw and Th represent the width and height of the predicted bounding box; stride represents the downsampling factor; anchor_grid represents the product of the resolution of the feature map after downsampling and the stride.

[0055] In the loss function, the CIoU loss is calculated using the coordinates of the predicted bounding box and the actual target bounding box, which can be expressed as:

[0056]

[0057] Where, ρ 2 (β,β gt ) represents the square of the Euclidean distance between the center points of the predicted box and the ground truth box; c represents the diagonal of the smallest bounding rectangle of the predicted box and the ground truth box; α is a weighting coefficient used to balance distance-related terms and aspect ratio-related terms; IoU represents the area of ​​the intersection of the predicted box and the ground truth box divided by the area of ​​their union.

[0058] In the loss function, the classification loss is calculated using cross-entropy loss. For each predicted bounding box, if the predicted bounding box is assigned to a true object, the cross-entropy loss between the predicted probability distribution of the bounding box and the true class can be expressed as:

[0059]

[0060] Where n represents the total number of categories, y represents the true category, and the predicted category probability distribution is p = (p1, ..., p2) / (p3, ..., p4) / (p5, ..., p6) / (p7) / (p8) / (p9) / (p1, ..., p1) / (p1) / (p1) / (p2 ...

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

[0062] In the loss function, the target confidence loss uses binary cross-entropy loss. Its calculation method is as follows:

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

[0064] The final loss function is a 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] Where, λ cls , λ box , λ obj These are the weighting coefficients for classification loss, bounding box loss, and target confidence loss, used to adjust the relative importance of each loss component in the total loss.

[0067] Another aspect of the present invention provides a remote sensing target detection method apparatus based on 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 apparatus based on multi-scale and efficient attention mechanism performs the remote sensing target detection method based on 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 aforementioned remote sensing target detection method based on a multi-scale and efficient attention mechanism.

[0069] In summary, this invention introduces MIA-S, MIA-M, and MIA-L modules based on multi-scale efficient attention into the target detection model. This allows the model to use different receptive fields for targets of different sizes, effectively addressing the scale effect problem in remote sensing images. Replacing ordinary convolutions in YOLOv7 with depthwise separable convolutions not only reduces computation but also improves the model's generalization ability, enabling it to more accurately identify targets in complex and varied remote sensing image backgrounds. Although an attention mechanism is introduced, this invention reduces computational complexity by designing an efficient attention IAEF module, allowing the model to meet real-time detection requirements while maintaining high accuracy. The use of depthwise separable convolutions further reduces the model's computational load and improves inference speed. The introduction of multi-scale efficient attention modules allows the model to better cope with the problem of high similarity between targets and backgrounds in remote sensing images and susceptibility to interference from similar backgrounds, improving the model's robustness. By fusing features of different scales in the neck network, the model can capture richer information, further enhancing its recognition ability. This invention significantly improves the accuracy and real-time performance of target detection in remote sensing images by introducing multi-scale and efficient attention mechanisms, combined with techniques such as depthwise separable convolution and reparameterized convolution, thereby enhancing the model's robustness and adaptability. This achievement provides a new solution and technical approach for the field of remote sensing image target detection, possessing significant theoretical and practical value.

[0070] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, 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. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0071] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications or substitutions should be covered within 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: A remote sensing image of the target is acquired, and the image is input into a target detection model to obtain the target detection result. The target detection model includes: replacing the ordinary 3×3 convolution in the 6th layer of the YOLOv7 network ELANBlock with a 3×3 depthwise separable convolution; replacing the ordinary 3×3 convolution in the 8th layer of the YOLOv7 network ELANBlock with a 5×5 depthwise separable convolution; replacing the ordinary 3×3 convolution in the 10th layer of the YOLOv7 network ELANBlock with a 7×7 depthwise separable convolution; replacing the 17th and 23rd layers of the YOLOv7 network ELANBlock with the MIA-M module based on multi-scale efficient attention; replacing the 21st layer of the YOLOv7 network ELANBlock with the MIA-S module based on multi-scale efficient attention; and replacing the 25th layer of the YOLOv7 network ELANBlock with the MIA-L module based on multi-scale efficient attention, thus obtaining the target detection model. The MIA-S, MIA-M, and MIA-L modules are multi-scale efficient attention modules with progressively larger scales. Each module includes two 1×1 regular convolutions, three 3×3 regular convolutions, and an efficient attention IAEF module. The two 1×1 regular convolutions are used to extract features from the input feature map to obtain features y1 and y2, respectively. The efficient attention IAEF module performs efficient attention operations on feature y2 to obtain feature y3. The first 3×3 regular convolution is used to extract features from feature y3 to obtain feature y4. The second 3×3 regular convolution is used to extract features from feature y4 to obtain feature y5. The third 3×3 regular convolution is used to extract features from feature y5 to obtain feature y6. Features y1, y2, y3, y4, y5, and y6 are concatenated to obtain the output feature. The efficient attention IAEF module is used to perform efficient attention operations on feature y2 to obtain feature y3, including: in, and These are two learnable parameters; Indicates will Perform mean calculation along the channel dimension; This represents a 3×3 grouped convolution; Indicates the concatenation function; express Depth-separable convolutions.

2. The remote sensing target detection method based on multi-scale and efficient attention mechanism according to claim 1, characterized in that, The YOLOv7 network layers 8 and 10, ELANBlock, share the same network structure, both consisting of three 1×1 ordinary convolutions and four 3×3 ordinary convolutions. The first and second 1×1 ordinary convolutions extract features f1 and f2 from the input feature map, respectively. The first 3×3 ordinary convolution extracts feature x1 from feature f2, the second 3×3 ordinary convolution extracts feature x2 from feature x1, and the third 3×3 ordinary convolution extracts feature x3 from feature x2. The fourth 3×3 ordinary convolution extracts feature x4 from feature x3. Features f1, f2, x2, and x4 are concatenated to obtain feature f3, which is then 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 convolutions in the 6th layer of YOLOv7 ELANBlock with 3×3 depthwise separable convolutions includes replacing the first, third, and fourth ordinary convolutions in the 6th layer of YOLOv7 ELANBlock with 3×3 depthwise separable convolutions.

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

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

6. The remote sensing target detection method based on multi-scale and efficient attention mechanism according to claim 1, characterized in that, In the MIA-S module The value of k in the MIA-M module is 3. The value of k in the MIA-L module is 5. The value of k in the equation is 7.

7. A remote sensing target detection device based on a multi-scale and efficient attention mechanism, characterized in that, It includes a processor and a memory; the memory is used to store computer programs; the processor is connected to the memory and is used to execute the computer programs stored in the memory, so that the remote sensing target detection device based on multi-scale and efficient attention mechanism performs the remote sensing target detection method based on multi-scale and efficient attention mechanism according to any one of claims 1 to 6.

8. A computer-readable storage medium storing a program, characterized in that, When the program is executed by the processor, it implements the remote sensing target detection method based on multi-scale and efficient attention mechanism as described in any one of claims 1-6.

Images