A remote sensing small target detection method based on large kernel convolution

Abstract

The application provides a remote sensing small target detection method based on large kernel convolution, and relates to the technical field of remote sensing image target detection. The method designs a C2f_UniRepLKNetBlock module, which is used in the backbone network. The module can flexibly adjust the convolution task according to the specific convolution kernel size of each convolution operation, flexibly aggregate multi-scale information, fully extract relevant feature information, play the advantages of large convolution kernel and effectively improve the detection precision. An LSKA-SPPF module is designed as a spatial pyramid pooling module of the backbone network. The LSKA-SPPF module combines the advantages of the self-attention mechanism and the large kernel convolution, decomposes the two-dimensional convolution into one-dimensional convolution, thereby reducing the interference of background information, reducing the computational complexity and memory occupation. In the regression loss calculation, the NWD method is introduced, the traditional CIoU method is combined with the NWD method, the sensitivity for detecting different scale objects in the remote sensing image is reduced, and the detection capability is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of remote sensing image target detection, and particularly relates to a remote sensing small target detection method based on large kernel convolution. BACKGROUND

[0002] The target detection task refers to finding all the targets or objects of interest in an image, determining their categories and positions, and is one of the core problems in the field of computer vision. Due to different appearances and shapes of various objects, and the influence of factors such as illumination and occlusion during imaging, target detection has always been a very challenging problem, and has attracted the attention of many researchers. Its application is also very extensive. Target detection has been applied in various fields, mainly including aerospace, autonomous driving cars, video surveillance and the like. Remote sensing small target detection is a timely detection of remote sensing images collected by remote sensing devices, is a long-distance and high real-time detection task, and is a special target detection task.

[0003] Traditional target detection methods are generally applicable to most detection tasks and have good effects on large and medium-sized detection. However, for remote sensing small target images, due to the mutual occlusion and small size of the objects, there are often omissions during feature extraction, resulting in poor detection results and causing false detection and missed detection problems. Therefore, a targeted and real-time method must be proposed to address the problems in target detection tasks of remote sensing images, so as to achieve efficient detection of remote sensing images. For example, the Chinese patent "CN118918487A Remote Sensing Image Target Detection Method Based on MCD-YOLO Network" extracts information of different scales and orientations, and improves the detection effect. This patent takes Yolov8 as the basic model, and proposes a C2f-LVMB module in the backbone network that can obtain different direction information and different scale information. The CA attention mechanism is introduced in the SPPF part to better capture feature map information, and the effective extraction of remote sensing image feature information is realized. However, it ignores another problem of remote sensing images: small targets account for a large proportion. This method does not consider the flexibility of feature extraction when extracting information of different scales and orientations, resulting in many small targets still being unable to be successfully detected. The Chinese patent "CN118781329A Small Target Detection Method for Ships in Optical Remote Sensing Images" improves the detection accuracy of remote sensing small targets by introducing a rotation loss and an attention mechanism. This patent takes Yolov5 as the basic model, introduces the CBAM attention mechanism in the model, and introduces the rotation angle in the bounding box regression loss, adjusts the horizontal detection box to the rotation detection box based on the field side representation method, and finally realizes the effective detection of small targets. This method increases the rotation loss function and the attention mechanism, greatly increasing the time complexity of the entire algorithm. In addition, the attention mechanism method used in this patent does not fully consider the long-distance dependence between different pixel points in remote sensing images, resulting in background information affecting the final detection results. SUMMARY

[0004] The technical problem to be solved by the present application is to address the shortcomings of the prior art and provide a remote sensing small target detection method based on large kernel convolution, which flexibly adjusts the convolution task, flexibly aggregates multi-scale information, fully extracts relevant feature information, and effectively improves the detection capability of small targets in remote sensing images.

[0005] To solve the above technical problems, the technical solution adopted by the present application is:

[0006] The present application provides a remote sensing small target detection method based on large kernel convolution, comprising the following steps:

[0007] S1, establish a remote sensing image dataset, preprocess remote sensing images in the remote sensing image dataset, and divide the preprocessed remote sensing image dataset into a training set, a validation set and a test set according to a proportion;

[0008] The method for preprocessing remote sensing images in the remote sensing image dataset is as follows: the remote sensing images in the remote sensing image dataset are cut into sub-images with a fixed size, the labels of the remote sensing images are cut to obtain sub-labels adapted to the sub-images, and the sub-images with the fixed size and the corresponding sub-labels are divided into the training set, the validation set and the test set according to the proportion;

[0009] S2, a UL-YOLOv8 remote sensing small target detection model based on large kernel convolution is established by taking YOLOv8 as a basic network, the C2f module in the backbone network of YOLOv8 is replaced by a C2f_UniRepLKNetBlock module, the SPPF spatial pyramid pooling module in the backbone network of YOLOv8 is replaced by an LSKA-SPPF spatial pyramid pooling module, and a NWD loss function is introduced in the loss function part in combination with the original CIOU loss function, so as to establish the UL-YOLOv8 remote sensing small target detection model based on large kernel convolution;

[0010] S2.1, based on the C2f module of the YOLOv8 model backbone network, a UniRepLKNetBlock module is introduced to replace the original Bottleneck module of the C2f module, a C2f_UniRepLKNetBlock module is constructed, and the C2f_UniRepLKNetBlock module is used to replace the C2f module in the YOLOv8 model backbone network;

[0011] Based on the C2f module of the YOLOv8 model backbone network, a UniRepLKNetBlock module is introduced to replace the original Bottleneck module of the C2f module, a C2f_UniRepLKNetBlock module is constructed, and for the input feature map of the C2f_UniRepLKNetBlock module, the channel number of the input feature map is adjusted to a pre-set output channel number by using a first 1×1 convolution, then the adjusted feature map is divided into two sub-branches by using a Split division function, the first sub-branch is processed by a plurality of UniRepLKNetBlock modules, the output results of the UniRepLKNetBlock modules are spliced with the second sub-branch in the channel dimension, and a second 1×1 convolution is used to adjust the feature map channel to a pre-set output channel number, so as to obtain the output feature map of the C2f_UniRepLKNetBlock module;

[0012] S2.2, based on the YOLOv8 model backbone network SPPF spatial pyramid pooling module, introduce LSKA attention mechanism module, build LSKA-SPPF module, use LSKA-SPPF module to replace SPPF spatial pyramid pooling module in YOLOv8 model backbone network;

[0013] Based on the YOLOv8 model backbone network SPPF spatial pyramid pooling module, introduce LSKA attention mechanism module, build LSKA-SPPF module, for the input feature map of LSKA-SPPF module, use the third 1×1 convolution to reduce dimension and feature fusion, to reduce the complexity of calculation; The output of the third 1×1 convolution is subjected to three times of maximum pooling, the pooling result of each time is spliced with the output of the third 1×1 convolution, and the spliced result is input into the LSKA attention mechanism module. The output channel number of the LSKA attention mechanism module is reduced to the specified output channel number by using the fourth 1×1 convolution to obtain the output of the LSKA-SPPF module;

[0014] LSKA attention mechanism decomposes a K×K large kernel convolution into an expansion convolution, a (2d-1)×(2d-1) depth convolution, and a 1×1 convolution, where d is the expansion rate, Downward rounding operation; then the (2d-1)×(2d-1) depth convolution is split into two cascaded one-dimensional separable kernels 1×(2d-1) and (2d-1)×1, and The expansion convolution of is split into two cascaded one-dimensional separable kernels and Convert the two-dimensional convolution to one-dimensional convolution;

[0015] S2.3, in the loss function part, introduce NWD loss function and original CIOU loss function, assign different weights to NWD loss function and CIOU loss function, reduce the sensitivity of loss function to small object position deviation:

[0016] NWD loss function performs two-dimensional Gaussian distribution on the real boundary box and the predicted boundary box, and then calculates the similarity through normalized Wasserstein distance, and the calculation formula is as follows:

[0017]

[0018] Where (cx1, cy1, w1, h1) and (cx2, cy2, w2, h2) correspond to the center horizontal and vertical coordinates, the predicted box width and height of the predicted boundary box and the real boundary box, cx i is the x-axis coordinate of the center point of the boundary box, cy i is the y-axis coordinate of the center point of the boundary box, wi is the width of the bounding box, h i is the height of the bounding box, N1, N2 represent the predicted bounding box and the real bounding box the result of the Gaussian distribution, w2 2 (N1, N2) represents the final normalized distance;

[0019] The NWD loss function is shown in the following formula:

[0020]

[0021] where C is the number of classes of the data set;

[0022] S3, using the preprocessed remote sensing image data set to train and verify the UL-YOLOv8 remote sensing small target detection model, setting the training related parameters, after the training is completed, saving the trained UL-YOLOv8 remote sensing small target detection model weight;

[0023] S4, using the trained UL-YOLOv8 remote sensing small target detection model weight to detect the small target contained in the remote sensing image, inputting the remote sensing image to be detected into the UL-YOLOv8 remote sensing small target detection model, obtaining the detected small target bounding box and label information;

[0024] The beneficial effects produced by the above technical scheme are that: the remote sensing small target detection method based on large kernel convolution provided by the application designs a C2f_UniRepLKNetBlock module, which is used in the backbone network. The module can flexibly adjust the convolution task according to the specific convolution kernel size of each convolution operation, flexibly aggregate multi-scale information, fully extract relevant feature information, play the advantages of large convolution kernel, and effectively improve the detection precision. The LSKA-SPPF module is designed as a spatial pyramid pooling module of the backbone network. The LSKA-SPPF module combines the advantages of self-attention mechanism and large kernel convolution, decomposes two-dimensional convolution into one-dimensional convolution, thereby reducing the interference of background information, reducing the computational complexity and memory occupation. In the regression loss calculation, the NWD method is introduced, the traditional CIoU method is combined with the NWD method, the sensitivity of different scale object detection in the remote sensing image is reduced, and the detection ability is improved. Compared with the prior art, the application can effectively solve the problems of overlapping and small target in remote sensing images, and greatly improve the problems of occlusion, overlapping and small target. BRIEF DESCRIPTION OF DRAWINGS

[0025] Figure 1 The UL-YOLOv8 remote sensing small target detection model structure diagram based on large kernel convolution provided by the embodiment of the application;

[0026] Figure 2 A UniRepLKNetBlock structure diagram provided for an embodiment of the present application;

[0027] Figure 3 A Dilated Reparam Block structure diagram provided for an embodiment of the present application;

[0028] Figure 4 An LSKA-SPPF module structure diagram provided for an embodiment of the present application;

[0029] Figure 5 A comparison diagram of detection results of a YOLOv8 model and an UL-YOLOv8 remote sensing small target detection model provided for an embodiment of the present application, wherein (a) is a remote sensing image to be detected, (b) is a detection result of detection using the YOLOv8 model, and (c) is a detection result of detection using the UL-YOLOv8 remote sensing small target detection model provided for the embodiment. DETAILED DESCRIPTION

[0030] The specific embodiments of the present application will be further described in detail below in combination with the drawings and embodiments. The following embodiments are used to illustrate the present application, but not to limit the scope of the present application.

[0031] A remote sensing small target detection method based on large kernel convolution provided in the embodiment, an UL-YOLOv8 remote sensing small target detection model based on large kernel convolution is established, as shown in Figure 1 The method includes the following steps:

[0032] S1, a remote sensing image dataset is established, and the remote sensing images in the remote sensing image dataset are preprocessed, and the preprocessed remote sensing image dataset is divided into a training set, a validation set and a test set according to a proportion;

[0033] The method for preprocessing the remote sensing images in the remote sensing image dataset is: cutting the remote sensing images in the remote sensing image dataset into sub-images with a fixed size, cutting the labels of the remote sensing images to obtain sub-labels adapted to the sub-images, and dividing the sub-images with the fixed size and the corresponding sub-labels into the training set, the validation set and the test set according to the proportion;

[0034] Since the scale of the original remote sensing images is too large and the sizes are different, in the embodiment, the remote sensing images and labels in the dataset are cut, and the remote sensing images in the dataset are cut into sub-images with a resolution of 1024*1024, and the cutting method can effectively collect more features of small targets in the remote sensing images;

[0035] S2, a UL-YOLOv8 remote sensing small target detection model based on large kernel convolution is established by taking YOLOv8 as a basic network, replacing the C2f module in the backbone network of YOLOv8 with a C2f_UniRepLKNetBlock module; replacing the SPPF spatial pyramid pooling module in the backbone network of YOLOv8 with a LSKA-SPPF spatial pyramid pooling module; introducing a NWD loss function in the loss function part combined with the original CIOU loss function, and establishing a UL-YOLOv8 remote sensing small target detection model based on large kernel convolution;

[0036] S2.1, based on the C2f module of the YOLOv8 model backbone network, the UniRepLKNetBlock module is introduced to replace the original Bottleneck module of the C2f module, and the C2f_UniRepLKNetBlock module is constructed, and the C2f_UniRepLKNetBlock module is used to replace the C2f module in the YOLOv8 model backbone network;

[0037] Based on the C2f module of the YOLOv8 model backbone network, the UniRepLKNetBlock module is introduced to replace the original Bottleneck module of the C2f module, and the C2f_UniRepLKNetBlock module is constructed, and for the input feature map of the C2f_UniRepLKNetBlock module, the channel number of the input feature map is adjusted to the pre-set output channel number by using the first 1×1 convolution, then the adjusted feature map is divided into two sub-branches by using the Split division function, wherein the first sub-branch is processed by multiple UniRepLKNetBlock modules, the output results of the multiple UniRepLKNetBlock modules are spliced with the second sub-branch in the channel dimension, and the feature map channel is adjusted to the pre-set output channel number by using the second 1×1 convolution, to obtain the output feature map of the C2f_UniRepLKNetBlock module;

[0038] The specific structure of the UniRepLKNetBlock module is shown in Figure 2 The UniRepLKNetBlock module includes a DW (depthwise) conv, an SE (Squeeze-and-Excitation) Block, an FFN feedforward network, and a BN batch normalization layer, wherein the SE Block structure can increase the depth of the UniRepLKNetBlock module and improve the fitting capability, the DW conv can flexibly change the convolution task according to the specific convolution kernel size, and the multiple selection of convolution methods makes the UniRepLKNetBlock module structure more flexible.

[0039] The DW conv includes two convolution modes of Dilated Reparam Block convolution and depthwise 3*3 convolution, and the DW conv selects to use the Dilated Reparam Block convolution or the depthwise 3*3 convolution mode for calculation according to the convolution task. When the convolution kernel size used in the convolution task is small, the DW conv uses depthwise convolution calculation, otherwise the DW conv uses Dilated Reparam Block calculation. The specific operation process of the Dilated Reparam Block convolution is as shown in Figure 3 As shown, by using dilated convolution in parallel to capture sparse features in the feature map, using multiple dilated convolutions with different dilation rates to jointly perform this convolution operation, without affecting the depth of the model, it provides a wider receptive field and enhances the advantages brought by the large kernel convolution. The feature map input into the Dilated Reparam Block is respectively convolved using convolution kernels with different scales and hole rates, and the BN operation is performed, the outputs obtained by using convolution kernels with different scales and hole rates for convolution operation and BN operation are superimposed to obtain the output of the Dilated Reparam Block; the Dilated Reparam Block can also be reparameterized into an equivalent large kernel convolution. The large kernel convolution after reparameterization can enhance the spatial information capturing capability of the UL-YOLOv8 remote sensing small target detection model while keeping the number of learnable parameters and computational efficiency;

[0040] The SE Block is a channel attention mechanism module, including the steps of extracting features, performing average pooling or maximum pooling on the feature map to obtain the weight of each channel, and multiplying each channel with its own weight, so that different channels have different forces on the result;

[0041] The FFN feedforward network is a linear transformation layer, which is used to complete the dimension transformation of input data to output data, including a first full connection layer, a Relu activation layer and a second full connection layer. The activation layer is used to increase the expression ability of the UniRepLKNetBlock module, so that the UniRepLKNetBlock module can capture complex features and patterns. The full connection layer is used to increase and decrease the dimension of the feature map, so as to obtain more rich and accurate feature representation;

[0042] The BN batch normalization layer makes the output feature map of each layer keep the same distribution, and alleviates the phenomenon of feature distribution deviation in the training process;

[0043] S2.2, SPPF spatial pyramid pooling module based on YOLOv8 model backbone network, introduce LSKA attention mechanism module, construct LSKA-SPPF module, use LSKA-SPPF module to replace SPPF spatial pyramid pooling module in YOLOv8 model backbone network;

[0044] The LSKA attention mechanism module is introduced into the SPPF spatial pyramid pooling module, and the LSKA-SPPF module is formed by fusing the SPPF spatial pyramid pooling module. The LSKA-SPPF module can efficiently decompose large convolution kernels while reducing time complexity and reducing computational consumption. The structure of the LSKA-SPPF module is as shown in Figure 4 ;

[0045] The SPPF spatial pyramid pooling module based on the YOLOv8 model backbone network introduces the LSKA attention mechanism module to construct the LSKA-SPPF module. For the input feature map of the LSKA-SPPF module, a first 1×1 convolution is used for dimension reduction and feature fusion to reduce the complexity of the calculation. The output of the third 1×1 convolution is subjected to three times of maximum pooling. The pooling result of each time is spliced with the output of the third 1×1 convolution. The spliced result is input into the LSKA attention mechanism module. The output channel number of the LSKA attention mechanism module is reduced to a specified output channel number by using a fourth 1×1 convolution to obtain the output of the LSKA-SPPF module.

[0046] The LSKA attention mechanism is based on the LAK attention mechanism and combines the advantages of self-attention mechanism and large kernel convolution. The large kernel convolution operation is decomposed to capture long-distance relationships. Specifically, a K×K large kernel convolution is decomposed into an expansion convolution, a (2d-1)×(2d-1) depth convolution, and a 1×1 convolution, where d is the expansion rate, downward rounding operation; the (2d-1)×(2d-1) depth convolution is split into two cascaded one-dimensional separable kernels 1×(2d-1) and (2d-1)×1, the expansion convolution of K×K is split into two cascaded one-dimensional separable kernels and The two-dimensional convolution is converted into a one-dimensional convolution, which effectively reduces the computational complexity and realizes the capture of long-distance information and the establishment of the correlation between different points, effectively avoiding the shortcomings in the channel dimension.

[0047] Large Separable Kernel Attention (LSKA) decomposes large kernel convolution operations to capture long-range relationships, capturing long-range dependencies between different points and establishing correlations between different points to generate self-attention feature maps, avoiding the shortcomings in the channel dimension. At the same time, the two-dimensional convolution kernel of the deep convolution layer is decomposed into a cascaded horizontal one-dimensional and vertical one-dimensional kernel, effectively reducing the problem of increased computational cost caused by large convolution kernel, and integrating it into the SPPF structure, which can enhance long-range dependencies and effectively extract information between different scales, thereby improving detection efficiency;

[0048] S2.3, in the loss function part, introduce NWD loss function combined with the original CIOU loss function, assign different weights to NWD loss function and CIOU loss function, reduce the sensitivity of loss function to small object position deviation:

[0049] In this embodiment, based on the loss function CIOU of YOLOv8, the NWD loss function is introduced, so that the NWD loss function and the CIOU loss function are calculated by loss according to the proportion of 1:1, which can reduce the sensitivity to small object position deviation, and effectively improve the detection result of small target in remote sensing image;

[0050] The NWD loss function performs two-dimensional Gaussian distribution on the real bounding box and the predicted bounding box, and then calculates the similarity by normalized Wasserstein distance, and the calculation formula is as follows:

[0051]

[0052] Where (cx1, cy1, w1, h1) and (cx2, cy2, w2, h2) correspond to the center horizontal and vertical coordinates, the predicted box width and height of the predicted bounding box and the real bounding box, respectively, cx i is the x-axis coordinate of the center point of the bounding box, cy i is the y-axis coordinate of the center point of the bounding box, w i is the width of the bounding box, h i is the height of the bounding box, N1 and N2 represent the predicted bounding box and the real bounding box Gaussian distribution result, indicates the final normalized distance;

[0053] The NWD loss function is shown in the following formula:

[0054]

[0055] Where C is the number of classes in the dataset;

[0056] The advantage of calculating the distance using the Gaussian distribution is that the similarity can be calculated even if there is no overlap between the two bounding boxes, and the method has smoothness to the position deviation and is not sensitive to the position deviation of small objects, so it is suitable for detecting small objects in remote sensing images.

[0057] S3, using the preprocessed remote sensing image dataset to train and verify the UL-YOLOv8 remote sensing small target detection model, setting the training related parameters, after the training is completed, saving the trained UL-YOOLv8 remote sensing small target detection model weight;

[0058] In this embodiment, the dataset is divided into training set, validation set and test set according to 7:1:2, the training coefficient is set to 300 rounds, and 8 pictures are input each time, after the training is completed, the trained weight is saved for detecting small targets contained in remote sensing images.

[0059] S4, using the trained UL-YOLOv8 remote sensing small target detection model weight to detect small targets contained in remote sensing images, inputting the remote sensing image to be detected into the UL-YOLOv8 remote sensing small target detection model to obtain the detected small target bounding box and label information;

[0060] The remote sensing small target detection method based on large kernel convolution provided in this embodiment uses large kernel convolution to flexibly adjust convolution tasks, flexibly aggregates multi-scale information, fully extracts relevant feature information, and effectively improves the model performance, which can effectively solve the problems of overlapping and more small targets in remote sensing images. This embodiment uses YOLOv8 model and UL-YOLOv8 remote sensing small target detection model to detect the remote sensing image to be detected, and the detection results are as shown in Figure 5 Figure 5 (a) is a remote sensing image to be detected, Figure 5 (b) is the detection result of using YOLOv8 model for detection, Figure 5 (c) is the detection result of using the UL-YOLOv8 remote sensing small target detection model provided in this embodiment for detection, and Figure 5 It can be known that compared with the traditional YOLOv8 model, the remote sensing small target detection method based on large kernel convolution provided in this embodiment has great improvement in dealing with the problems of occlusion and overlap, and can detect more small targets.

[0061] ​The remote sensing small target detection method based on large kernel convolution provided in the embodiment not only realizes the small target detection task of the remote sensing image, but also can play the advantage of the large kernel convolution according to the image features, flexibly adjust the convolution task, solve the problems of the traditional target detection method, such as small target missed detection, large difference between different target scales, strong background interference and the like in the remote sensing image, and greatly improve the precision and speed of the remote sensing target detection technology, thereby providing strong support for realizing the real-time and accurate detection of the remote sensing small target in the fields of military detection, intelligent transportation and the like.

[0062] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present application, but not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement for part or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope defined by the claims of the present application.

Claims

1. A remote sensing small target detection method based on large kernel convolution, characterized in that: The method comprises the following steps: S1, establishing a remote sensing image dataset, preprocessing remote sensing images in the remote sensing image dataset, and dividing the preprocessed remote sensing image dataset into a training set, a validation set and a test set according to a proportion; S2, using YOLOv8 as a basic network to establish a UL-YOLOv8 remote sensing small target detection model based on large kernel convolution, replacing the C2f module in the backbone network of YOLOv8 with a C2f_UniRepLKNetBlock module; replacing the SPPF spatial pyramid pooling module in the backbone network of YOLOv8 with a LSKA-SPPF spatial pyramid pooling module; introducing a NWD loss function in the loss function part in combination with the original CIOU loss function, and establishing a UL-YOLOv8 remote sensing small target detection model based on large kernel convolution; S3, training and verifying the UL-YOLOv8 remote sensing small target detection model using the preprocessed remote sensing image dataset, setting training related parameters, saving the trained UL-YOLOv8 remote sensing small target detection model weight after training is completed; S4, detecting small targets contained in a remote sensing image using the trained UL-YOLOv8 remote sensing small target detection model weight, inputting the remote sensing image to be detected into the UL-YOLOv8 remote sensing small target detection model to obtain a detected small target bounding box and label information; The S2 comprises: S2.1, based on the C2f module of the YOLOv8 model backbone network, introducing a UniRepLKNetBlock module to replace the original Bottleneck module of the C2f module, constructing a C2f_UniRepLKNetBlock module, and using the C2f_UniRepLKNetBlock module to replace the C2f module in the YOLOv8 model backbone network; S2.2, based on the SPPF spatial pyramid pooling module of the YOLOv8 model backbone network, introducing a LSKA attention mechanism module to construct a LSKA-SPPF module, and using the LSKA-SPPF module to replace the SPPF spatial pyramid pooling module in the YOLOv8 model backbone network; S2.3, in the loss function part, introducing a NWD loss function in combination with the original CIOU loss function, assigning different weights to the NWD loss function and the CIOU loss function, and reducing the sensitivity of the loss function to the position deviation of the micro object; The NWD loss function calculates the similarity by performing two-dimensional Gaussian distribution on the real bounding box and the predicted bounding box after normalization, and the calculation formula is as follows: wherein, and correspond to the horizontal and vertical coordinates of the center of the predicted and ground truth bounding boxes, respectively, the width and height of the predicted bounding box, is the x-coordinate of the center point of the bounding box, is the y-coordinate of the center point of the bounding box, is the width of the bounding box, is the height of the bounding box, denotes the predicted bounding box and the ground truth bounding box the result of the Gaussian distribution, denotes the final normalized distance; The NWD loss function is shown in the following formula: Wherein, C is the number of classes of the dataset.

2. The remote sensing small target detection method based on large kernel convolution according to claim 1, characterized in that: The method for preprocessing the remote sensing images in the remote sensing image dataset in S1 is: cutting the remote sensing images in the remote sensing image dataset into sub-images with a fixed size, cutting the labels of the remote sensing images to obtain sub-labels adapted to the sub-images, and dividing the sub-images with the fixed size and their corresponding sub-labels into a training set, a validation set and a test set in proportion.

3. The remote sensing small target detection method based on large kernel convolution according to claim 1, characterized in that: The S2.1 specific method is: based on the C2f module of the YOLOv8 model backbone network, introducing a UniRepLKNetBlock module to replace the original Bottleneck module of the C2f module, constructing a C2f_UniRepLKNetBlock module, for the input feature map of the C2f_UniRepLKNetBlock module, adjusting the channel number of the input feature map to a pre-set output channel number by using a first 1*1 convolution, then using a Split segmentation function to divide the adjusted feature map into two sub-branches, wherein the first sub-branch is processed through a plurality of UniRepLKNetBlock modules, the results output by each UniRepLKNetBlock module are spliced with the second sub-branch in the channel dimension, and a second 1*1 convolution is used to adjust the feature map channel to a pre-set output channel number to obtain the output feature map of the C2f_UniRepLKNetBlock module.

4. The remote sensing small target detection method based on large kernel convolution according to claim 3, characterized in that: The S2.2 specific method is: based on the SPPF spatial pyramid pooling module of the YOLOv8 model backbone network, introducing a LSKA attention mechanism module to construct a LSKA-SPPF module, for the input feature map of the LSKA-SPPF module, using a third 1*1 convolution for dimension reduction and feature fusion to reduce the complexity of calculation; performing three maximum pooling operations on the output of the third 1*1 convolution, splicing the pooling results of each time with the output of the third 1*1 convolution, inputting the spliced result into the LSKA attention mechanism module, and using a fourth 1*1 convolution to reduce the output channel number of the LSKA attention mechanism module to a specified output channel number to obtain the output of the LSKA-SPPF module.

5. The large kernel convolution-based remote sensing small target detection method according to claim 4, characterized in that: The LSKA attention mechanism in S2.2 decomposes a KxK large kernel convolution into an dilated convolution of size (2d-1)x(2d-1), a depthwise convolution of size (2d-1)x(2d-1), and a 1x1 convolution, where d is the dilation rate, a floor operation, and decomposing the depthwise convolution of size (2d-1)x(2d-1) into two cascaded one-dimensional separable kernels 1x(2d-1) and (2d-1)x1, decomposing the dilated convolution of size (2d-1)x(2d-1) into two cascaded one-dimensional separable kernels and convert the two-dimensional convolution into one-dimensional convolution.

Images