Remote sensing image small target detection method based on improved YOLO deep learning

Abstract

The application discloses a kind of remote sensing image small target detection methods based on improved YOLO deep learning, on the deep learning network framework of YOLOV3, make improvement to convolution kernel, increase receptive field to distinguish target and background, adopt inverse residual structure to construct feature extraction module LKG Bottleneck, and adopt CSPNet to construct the main network of LKGNet, finally introduce GIou loss function to enhance the sensitivity of network to boundary, realize the small target detection of remote sensing image;In addition, on the improved LKGNet network, further introduce dual attention module, utilize channel attention to important channel weighting, utilize spatial attention to suppress background noise, and make improvement on the original feature pyramid network structure, introduce weighted bidirectional feature pyramid.The application has very outstanding performance in computing cost, small target detection precision, overall average precision and robustness to noise.

Description

Technical Field

[0001] This invention belongs to the field of remote sensing image classification technology, specifically relating to a method for detecting small targets in remote sensing images based on improved YOLO deep learning. Background Technology

[0002] With the rapid development of human technology, satellite sensor technology has made leaps and bounds, effectively compensating for the limitations of human means of exploring and observing the Earth, and acquiring a wide variety of remote sensing data. SAR (Synthetic Aperture Radar) satellite imagery: SAR imagery is based on the aperture principle, an active data acquisition method unaffected by weather conditions such as clouds and rain, enabling all-weather data acquisition while preserving excellent target texture information; Optical remote sensing imagery: Utilizing solar wavelengths reflected from ground objects for imaging, these images possess rich spectral and texture information, and can serve as datasets for deep learning.

[0003] In the field of remote sensing, the main task of target detection is to identify the target based on given remote sensing images, including information such as the target's category, location, and confidence level. This enables rapid target localization, detection, and identification, which has significant practical value in both civilian and military fields, such as specific feature detection, geographic information collection, urban vehicle detection, and military target reconnaissance. Early on, due to technological limitations, the quality and quantity of remote sensing images were insufficient, such as low spatial resolution, making it impossible to detect and identify individual man-made or natural objects. Only regional classification of images was possible. The field of target detection in remote sensing images started relatively late. Therefore, it is necessary to develop new and more effective feature extraction and classification methods to address the urgent needs of industry applications and the challenges of hyperspectral remote sensing image classification.

[0004] With advancements in science and technology, more and more high-resolution satellites are being launched, gradually enabling the acquisition of remote sensing images with meter-level or even sub-meter-level resolution. This provides us with detailed spatial location attributes and contour texture information of ground features, allowing us to directly identify individual objects from the images. Because the sensors in optical remote sensing satellite images are too far from ground objects, some objects may occupy only a few to tens of pixels in the image. Although these objects occupy a very small portion of the image, they can be of great value (such as detecting vehicles in cities or tanks in battlefield environments). Therefore, detecting small targets in optical remote sensing images is a highly valuable research direction.

[0005] There are two ways to define small targets: based on relative size and based on absolute size. The definition based on relative size is provided by the international organization SPIE, which defines targets smaller than 0.12% of the image size as small targets. The other definition, based on absolute size and provided by the MS COCO dataset, considers targets smaller than 32×32 pixels to be small targets.

[0006] In recent years, due to the rapid development of deep learning, this technology has demonstrated extraordinary performance, especially in big data analysis and processing. Deep learning has extremely deep network structures and a large number of learnable parameters, enabling it to accurately extract effective feature information from data and achieve accurate and rapid data processing. Therefore, deep learning technology has been widely used in the field of small target detection in remote sensing images.

[0007] Early traditional object detection algorithms lacked targeted strategies for generating candidate regions on images, resulting in redundant candidate regions and high computational costs. Furthermore, the need for manual selection of appropriate feature extraction operators for different datasets led to poor generalization, requiring constant adjustments to feature extraction operators for different datasets or target scenes. In 2012, AlexNet emerged, winning the image classification competition that year and spurring the rapid development of deep learning. Since then, deep learning methods have become dominant in image processing, primarily employing the following approaches:

[0008] Two-stage object detection algorithms generate candidate regions using a Region Proposal Network (RPN). While selecting candidate regions, the algorithm makes an initial prediction of the target's location. Then, it further refines and judges the target's location and category, obtaining the final result after two stages. Single-stage object detection algorithms do not generate candidate regions; the target's location and category information are directly regressed through the network to obtain the result. However, current two-stage object detection algorithms still have shortcomings in computational efficiency, classification accuracy, handling the specific characteristics of small targets in remote sensing images, and feature information fusion. Especially in the field of small target detection in remote sensing images, small target information is easily lost, leading to low detection accuracy. Summary of the Invention

[0009] Technical problem solved: This invention proposes a method for detecting small targets in remote sensing images based on improved YOLO deep learning, which has significant improvements in computational efficiency, classification accuracy, the specific characteristics of small targets in remote sensing images, and the weakness of feature information fusion.

[0010] Technical solution:

[0011] A method for detecting small targets in remote sensing images based on improved YOLO deep learning, the method comprising the following steps:

[0012] S1, Improve the original YOLOv3 network to generate the LKGNet-YOLO network; The LKGNet-YOLO network includes a Ghost module, an LKG Bottleneck module, an FPN module, and a detector;

[0013] The LKG Bottleneck module comprises a first convolutional layer, a first cross-stage network module (first LKGCSP), a second convolutional layer, a second cross-stage network module (second LKGCSP), a third convolutional layer, a third cross-stage network module (third LKGCSP), a fourth convolutional layer, and a fourth cross-stage network module (fourth LKGCSP) connected in sequence. The convolutional kernels of the first, second, third, and fourth convolutional layers are all 2*2, with a stride of 2, and are composed of a Conv layer, a BatchNormalization layer, and a SiLU layer connected in sequence.

[0014] The FPN module includes a fifth cross-stage network module (fifth LKGCSP), a sixth cross-stage network module (sixth cross-stage network module (sixth LKGCSP)), and a seventh cross-stage network module (seventh cross-stage network module (seventh LKGCSP)). A conv layer, an upsampling layer, and a connection layer are sequentially connected between the fifth and sixth cross-stage network modules (fifth LKGCSP) and between the sixth and seventh cross-stage network modules (sixth LKGCSP). Furthermore, a fourth cross-stage network module (fourth LKGCS) is also included. The output of P is connected to the input of the fifth cross-stage network module (fifth LKGCSP), the output of the third cross-stage network module (third LKGCSP) is connected to the input of the sixth cross-stage network module (sixth LKGCSP), and the output of the second cross-stage network module (second LKGCSP) is connected to the input of the seventh cross-stage network module (seventh LKGCSP). The outputs of the fifth cross-stage network module (fifth LKGCSP), the sixth cross-stage network module (sixth LKGCSP), and the seventh cross-stage network module (seventh LKGCSP) are each connected to the detector through a conv layer.

[0015] The number of blocks in the first cross-stage network module (first LKGCSP), the second cross-stage network module (second LKGCSP), the third cross-stage network module (third LKGCSP), the fourth cross-stage network module (fourth LKGCSP), the fifth cross-stage network module (fifth LKGCSP), the sixth cross-stage network module (sixth LKGCSP), and the seventh cross-stage network module (seventh LKGCSP) are 3, 3, 9, 3, 3, 3, and 3, respectively. Each of them includes two 7*7 convolutional layers and one 1*1 convolutional layer connected in sequence. First, the low-dimensional features of the input image are mapped to a high-dimensional space through two 7*7 convolutional layers for convolutional extraction, and then the dimensionality is reduced and output through the 1*1 convolutional layer.

[0016] S2, preprocessing the remote sensing images in the dataset to enhance the dataset, which contains an average of multiple small target objects in the remote sensing images in the dataset;

[0017] S3. The preprocessed data is imported into the improved YOLOv3 network for training.

[0018] Furthermore, in step S2, the preprocessing operations on the remote sensing images in the dataset include: random scale transformation, random shifting, random left and right flipping Mosaic, and color dithering algorithm preprocessing.

[0019] Furthermore, in the Ghost module, m original feature maps are generated using a single convolution. The operation formula Y′=X*f′ is used to optimize the f′ convolution kernel; where Let f' represent the convolution kernel, with parameters set the same as those in a regular convolution kernel. Then, the following formula is used to perform a less computationally intensive convolution on the features in Y', generating s feature maps resulting from convolution operations on the features generated in the previous step:

[0020]

[0021] Furthermore, in step S3, for the improved YOLOv3 network, the following formula is used for node feature fusion:

[0022]

[0023] In the formula, represents the convolution operation; Re represents upsampling or downsampling of the feature map; Indicates from node P m，n To node P p，q The weights of network connections are ensured to be greater than 0 using the ReLU function; ∈ 10. -4 This is to avoid numerical instability by adding a number close to 0, P. i，jLet i represent the characteristics of the node in the i-th row and j-th column, where i = 1, 2; j = 1, 2.

[0024] Furthermore, the remote sensing image small target detection method also includes the following steps:

[0025] A DA layer is introduced into the LKGCSP layer to generate the DALKGCSP layer. The DALKGCSP layer consists of a first 1*1 conv layer, a DA layer, n LKGB layers, a concat layer, and a second 1*1 conv layer connected in sequence. The input of the first conv layer is also connected to one of the inputs of the concat layer through a third 1*1 conv layer. The value of n is the same as the number of blocks in the corresponding DALKGCSP layer. The DA layer consists of a channel attention module and a spatial attention module connected in sequence.

[0026] Each LKGB layer consists of two 7*7 convolutional layers and one 1*1 convolutional layer connected in sequence. First, the low-dimensional features of the input image are mapped to a high-dimensional space through the two 7*7 convolutional layers for convolution extraction, and then the 1*1 convolutional layer is used to reduce the dimensionality and output the result.

[0027] Furthermore, the remote sensing image small target detection method also includes the following steps:

[0028] The eighth cross-stage network module (eighth LKGCSP) and the ninth cross-stage network module (ninth LKGCSP) are introduced into the FPN module to generate the BiFPN module;

[0029] The output of the third cross-stage network module (third LKGCSP) is connected through one of the inputs of the first linear fusion layer. The output of the seventh cross-stage network module (seventh LKGCSP) is directly input to the detector on one hand, and connected to the other input of the first linear fusion layer through the conv layer on the other hand. The first linear fusion layer performs linear fusion on the input data and then inputs it to the eighth cross-stage network module (eighth LKGCSP), which in turn inputs it to the detector.

[0030] The output of the fourth cross-stage network module (fourth LKGCSP) is connected to one of the inputs of the second linear fusion layer. The output of the eighth cross-stage network module (eighth LKGCSP) is connected to the other input of the second linear fusion layer through the conv layer. The second linear fusion layer performs linear fusion on the input data and then inputs it to the ninth cross-stage network module (ninth LKGCSP), which in turn inputs it to the detector.

[0031] Beneficial effects:

[0032] This invention presents a remote sensing image small target detection method based on improved YOLO deep learning. Addressing the problem of small targets occupying a small proportion of the image and being difficult to extract features from, it proposes the LKGNet-YOLO network. Previous studies neglected the impact of the receptive field on small target detection. By constructing new convolutional layers, the network's receptive field is expanded, achieving better small target detection results compared to other classic target detection algorithms, while also having a greater advantage in parameter reduction. Furthermore, to address the issues of a larger receptive field introducing additional background noise and the network's weak multi-layer feature fusion ability, the DA-LKGNet-YOLO model is further proposed. Through the use of a dual attention mechanism module and a bidirectional feature pyramid structure, it achieves better noise suppression and multi-layer feature information fusion effects, demonstrating a greater advantage in small target detection compared to other comparative networks. Attached Figure Description

[0033] Figure 1 This is a flowchart illustrating the method for detecting small targets in remote sensing images based on an improved YOLO deep learning architecture in an embodiment of the present invention.

[0034] Figure 2 This is a flowchart of the YOLO algorithm in an embodiment of the present invention;

[0035] Figure 3 This is a schematic diagram of Ghostmodule convolution in an embodiment of the present invention;

[0036] Figure 4 These are the performance comparison results of various networks in the embodiments of this invention;

[0037] Figure 5 This is a structural diagram of LKGNet-YOLO in an embodiment of the present invention;

[0038] Figure 6 This is a trend chart of the loss function of the LKGNet-YOLO network in an embodiment of the present invention;

[0039] Figure 7 The test set PR curve structure diagrams in this embodiment of the invention are as follows: (a) base network PR curve, (b) base+LKGNet network PR curve, (c) base+GIoU network PR curve, and (d) base+LKGNet+GIoU network PR curve.

[0040] Figure 8 This is a structural diagram of the dual attention mechanism (DA) module proposed in this embodiment of the invention;

[0041] Figure 9 These are the following diagrams presented in the embodiments of the present invention: (a) DA-LKGCSP structure diagram; (b) DA-LKGNet network structure diagram;

[0042] Figure 10 This is a schematic diagram of the DA-LKGNet-YOLO structure in an embodiment of the present invention;

[0043] Figure 11 This is a trend chart of the loss function of the DA-LKGNet-YOLO network in an embodiment of the present invention;

[0044] Figure 12 The test set PR curves in this embodiment of the invention are: (a) base network PR curve, (b) base+DA network PR curve, (c) base+BiFPN network PR curve, and (d) base+DA+BiFPN network PR curve. Detailed Implementation

[0045] The following embodiments are provided to enable those skilled in the art to more fully understand the present invention, but do not limit the invention in any way.

[0046] See Figure 1 This invention discloses a method for detecting small targets in remote sensing images based on improved YOLO deep learning, used to detect small targets in images from the UCAS-AOD v1.0 dataset, including the following steps:

[0047] Step A. Obtain the UCAS-AOD v1.0 dataset, which contains 910 images in two categories: cars and airplanes. There are 310 images of cars, with an average of 14.43 car targets per image and 600 images of airplanes, with an average of 5.98 airplane targets per image. These will serve as supporting data for the experiment. Then proceed to Step B.

[0048] Step B. Preprocess the data obtained in Step A by random scaling, random shifting, random left / right flipping, Mosaic, and color dithering algorithm (HSV-space); Mosaic is based on CutMix and uses four images for data augmentation. CutMix is ​​based on the formula... and Calculate the labels of the new training samples using r x ~Unif(0, W), and ry~Unif(0,H), Sampling is performed to augment the dataset; then proceed to step C. In step B, the dataset undergoes preprocessing such as random scaling, random shifting, random left and right flipping (Mosaic), and color dithering (HSV-space) to enhance its reliability. The dataset is then fed into Darknet-53 for extraction. A skipping method is used to construct residual convolutions and perform batch regularization. The K-means method is used to cluster the dataset and accelerate convergence.

[0049] Step C. Input the dataset into the Ghost module structure, via... The formula takes a specific feature value as input and uses the formula Y = X × f + b to calculate an output feature map with n channels. Ordinary convolution can be broken down into two steps, using one convolution to generate m original feature maps. The calculation formula Y′=X*f′ is used to optimize the f′ convolution kernel. The convolution kernel is defined as f′. To ensure consistent feature size after generation, the parameters of f′ are set the same as those in a regular convolution kernel. However, the resulting feature map has a relatively small number of channels. To obtain more channels, a formula is proposed for the features in Y′. Perform convolutions with less computation to generate s feature maps that are convolved with the features generated in the previous step; then proceed to step D.

[0050] Step D. Enter the Bottleneck structure. In the first layer, the LKG module is used to extract features and map low-dimensional features to a high-dimensional space. The original 1×1 and 3×3 convolutions in the first step are replaced with 7×7 convolutions, increasing the receptive field coverage and sensing more surrounding information. The feature extraction process is changed to mapping low-dimensional features to a high-dimensional space for convolution extraction. The LKG module has two 7×7 convolutions, which can effectively extract features. Then, in the second layer, 1×1 convolutions are used for feature extraction again. Finally, 1×1 convolutions are used to reduce the number of channels in the output features, and the common ReLU activation function is replaced with the GELU function. The LKG Bottleneck (Large Kernel Ghostmodule Bottleneck) is embedded into the CSP structure through dimensionality reduction using 1×1 convolutions, and the problem of increased parameters caused by large convolutional kernels is solved through cross-stage local networks.

[0051] In addition, the convolutional layers with a kernel of 3*3 and a stride of 2 used in the original spatial downsampling were modified to convolutional layers with a kernel of 2*2 and a stride of 2 to achieve downsampling. The number of blocks was adjusted to 3, 3, 9, and 3 to build the Ghostmodule structure for feature extraction.

[0052] Preferably, inspired by attention mechanisms and the weighted bidirectional feature pyramid (BiFPN) structure, a dual attention structure (DA) is proposed by concatenating channel attention and spatial attention mechanisms. This DA is introduced into the modified inverted residual structure after a 1*1 convolution, followed by an LKG Bottleneck structure, and Conv+Batch Normalization+SiLU is used as the ordinary convolutional layer to complete the inverted residual structure. The modified DA-LKGNet network then replaces the original YOLOv3 backbone network, and BiFPN replaces the original FPN structure. Mathematically, it is assumed that P... i，j Let the feature of the node in the i-th row and j-th column be represented, then using Feature fusion of nodes is performed. A DA-LKGNet-YOLO model is constructed for feature extraction and bounding box detection.

[0053] The performance results of the aforementioned small target detection method are illustrated below through an example.

[0054] I. Experimental Instructions

[0055] (1) Data selection and preprocessing: The UCAS-AOD v1.0 dataset was obtained, which contains 910 images in two categories: cars and airplanes. Among them, there are 310 images of cars, with an average of 14.43 car targets per image and 600 images of airplanes, with an average of 5.98 airplane targets per image. As experimental support data, the obtained data was preprocessed by random scaling, random shifting, random left and right flipping, Mosaic and color dithering algorithms (HSV-space) to enhance the dataset.

[0056] (2) Comparison Method

[0057] The validation of the model effectiveness is mainly divided into the validation of the remote sensing image small target detection algorithm based on GhostNet and cross-stage local networks; and the validation of the remote sensing image small target detection algorithm that integrates attention mechanism and weighted bidirectional feature pyramid.

[0058] (3) Evaluation indicators

[0059] The target detection results are analyzed through statistical and comparative accuracy mPA, feature map visualization, comparison of detection time, parameter quantity and model weight size of different models, complexity analysis, generalization analysis, etc.

[0060] II. Experimental Results

[0061] (1) The YOLOv3 object detection algorithm continues the ideas of the previous YOLO series, and its algorithm flowchart is as follows: Figure 2Based on this, the Ghost module was designed and extracted, and its module diagram is shown below. Figure 3 When the learning rate is 0.01, compared with other learning rates, the network can find the optimal solution faster and achieve the best parameter optimization. Therefore, 0.01 is selected as the initial learning rate for the LKGNet-YOLO model.

[0062] The best performance on the dataset was achieved using two 7×7 convolutional kernels, confirming that larger kernels have a larger receptive field, thus enabling better separation of the target from the background. However, excessively large kernels introduce too much background noise, preventing the network from accurately identifying small targets. Therefore, accuracy decreased with 9×9 kernels. This study ultimately replaced the 1×1 and 3×3 convolutions in Ghostmodule with two 7×7 convolutions. Comparison results with various network frameworks are shown below. Figure 4 By analyzing the data and selecting the Cross-Stage Local Network (CSPNet), the original YOLOv3 backbone network Darknet-53 was replaced with LKGNet, resulting in the LKGNet-YOLO network of this study. Figure 5 The network's detection accuracy improved by 2.7 percentage points, demonstrating that the network's large receptive field can improve the detection performance of small targets. Replacing the original loss function with the GIoU loss function improved mAP by 1.5 percentage points, as the GIoU function can better perceive the location of the target box, proving the effectiveness of the improved loss function. The results are as follows... Figure 6 , Figure 7 As shown.

[0063] The network does not effectively focus on the small target itself, but instead focuses on too much background area, which leads to unsatisfactory small target detection results. A comparative experiment was conducted, with IoU set to 0.5. The mAP and AP results of each detection algorithm are shown in Table 1. Compared with other single-stage detectors such as YOLOv3, SSD, FSSD, RFBNet, and RefineDet, LKGNet-YOLO achieved better performance in both Airplanne and Car target detection, with an mAP of 0.941 across all categories.

[0064] Table 1 Comparison of models in the UCAS-AOD dataset

[0065]

[0066]

[0067] The LKGNet-YOLO object detection algorithm proposed in this embodiment has slightly more parameters than YOLOv3-SPP and SSD compared to other object detection algorithms, but significantly fewer parameters than other object detection algorithms. In terms of detection speed, LKGNet-YOLO is comparable to YOLOv3-SPP and outperforms other object detection algorithms. In comparison of generalization analysis, the network shows good performance even on a small proportion of the training set. The experimental comparisons are shown in Tables 2 and 3.

[0068] Table 2 shows the detection time, number of parameters, and model weights for each model.

[0069]

[0070] Table 3. Experimental results under different proportions of data.

[0071]

[0072] Inspired by attention mechanisms and the weighted bidirectional feature pyramid (BiFPN) structure, a dual attention structure (DA) is proposed by concatenating channel attention and spatial attention mechanisms. Figure 8 As shown. The modified inverted residual structure is introduced into a 1*1 convolution, followed by an LKG Bottleneck structure, and then Conv+Batch Normalization+SiLU is used as the ordinary convolutional layer to complete the inverted residual structure, as shown. Figure 9 As shown: (a) DA-LKGCSP structure diagram; (b) DA-LKGNet network structure diagram. The original YOLOv3 backbone network is replaced with the modified DA-LKGNet network, and the original FPN structure is replaced with BiFPN. Mathematically, it is assumed that P... i，j Let the feature of the node in the i-th row and j-th column be represented, then using Feature fusion of nodes is performed. A DA-LKGNet-YOLO model is constructed, and data extraction is performed, as illustrated in the diagram below. Figure 10 . Figure 11This is a trend chart of the loss function of the DA-LKGNet-YOLO network in this embodiment of the invention. After 200 epochs, the loss function converges. Therefore, in this embodiment, the network model training epoch is set to 200 for ablation experiments. According to the PR curve, compared to object detection networks without an attention mechanism, the use of the DA module can improve the mAP of small targets, indicating that the spatial attention mechanism can concentrate the location of network-learned features, suppress the influence of surrounding noise on the network, and even small targets in background noise can be well identified, improving detection accuracy. A quantitative analysis of the detection performance of the DA-LKGNet-YOLO algorithm on the UCAS-AOD dataset is also presented. (See [link to relevant documentation]). Figure 12 This paper also introduces RFBNet, RetinaNet, Faster R-CNN, etc., and compares them with DA-LKGNet-YOLO. Table 4 shows the results of DA-LKGNet-YOLO and various detection methods at IoU=0.5. The DA-LKGNet-YOLO algorithm extracted in this design improves the mAP to 0.962 in the final detection results, which is 2.1 percentage points higher than the previous network performance. DA-LKGNet-YOLO achieved the best detection results for the Car category.

[0073] Table 4 Comparison results of various models in the UCAS-AOD dataset

[0074]

[0075] The LKGNet-YOLO object detection algorithm proposed in this embodiment has slightly more parameters than YOLOv3-SPP and SSD compared to other object detection algorithms, but significantly fewer parameters than other object detection algorithms. In terms of detection speed, LKGNet-YOLO is comparable to YOLOv3-SPP and outperforms other object detection algorithms. In terms of generalization analysis, even on a small proportion of the training set, the network shows good performance. The experimental comparisons are shown in Tables 5 and 6.

[0076] Table 5 shows the detection time, number of parameters, and model weights for each model.

[0077]

[0078] Table 6. Experimental results under different proportions of data.

[0079]

[0080] This embodiment relates to a method for detecting small targets in remote sensing images based on improved YOLO deep learning. Taking small targets in remote sensing images as the research object, this embodiment studies the detection of small targets in remote sensing images using deep learning technology. The main purpose is to solve the problem that small target information is easily lost in the field of small target detection in remote sensing images, resulting in low detection accuracy. An improved remote sensing image small target detection algorithm based on YOLO is proposed. The design contributions are: (1) In response to the problem that small targets occupy a small proportion in the image and their features are difficult to extract, LKGNet-YOLO is proposed. In previous studies, the influence of the receptive field on small target detection was ignored. By constructing a new convolutional layer, the receptive field of the network is expanded, and better small target detection results are achieved compared with other classic target detection algorithms. At the same time, it has a greater advantage in terms of the number of parameters; (2) In response to the problem that a larger receptive field will introduce some background noise and the network's weak multi-layer feature fusion ability, the DA-LKGNet-YOLO model is proposed. By using a dual attention mechanism module and a bidirectional feature pyramid structure, better noise suppression and multi-layer feature information fusion effects are achieved. Compared with other comparison networks, it has a greater advantage in small target detection.

Claims

1. A method for detecting small targets in remote sensing images based on improved YOLO deep learning, characterized in that, The method for detecting small targets in remote sensing images includes the following steps: S1, Improve the original YOLOv3 network to generate the LKGNet-YOLO network; The LKGNet-YOLO network includes a Ghost module, an LKG Bottleneck module, an FPN module, and a detector; The LKG Bottleneck module comprises a first convolutional layer, a first LKGCSP layer, a second convolutional layer, a second LKGCSP layer, a third convolutional layer, a third LKGCSP layer, a fourth convolutional layer, and a fourth LKGCSP layer connected in sequence; wherein, the convolutional kernels of the first, second, third, and fourth convolutional layers are all 2*2, and the stride is 2, and they are composed of a Conv layer, a BatchNormalization layer, and a SiLU layer connected in sequence; The FPN module includes a fifth LKGCSP layer, a sixth LKGCSP layer, and a seventh LKGCSP layer. A conv layer, an upsampling layer, and a connection layer are sequentially connected between the fifth and sixth LKGCSP layers, and between the sixth and seventh LKGCSP layers. The output of the fourth LKGCSP layer is connected to the input of the fifth LKGCSP layer, the output of the third LKGCSP layer is connected to the input of the sixth LKGCSP layer, and the output of the second LKGCSP layer is connected to the input of the seventh LKGCSP layer. The outputs of the fifth, sixth, and seventh LKGCSP layers are each connected to a detector via a conv layer. The number of blocks in the first, second, third, fourth, fifth, sixth, and seventh LKGCSP layers are 3, 3, 9, 3, 3, 3, and 3, respectively. Each layer includes two 7*7 convolutional layers and one 1*1 convolutional layer connected in sequence. The low-dimensional features of the input image are first mapped to a high-dimensional space through two 7*7 convolutional layers for convolution extraction, and then the dimensionality is reduced and output through the 1*1 convolutional layer. S2, preprocessing the remote sensing images in the dataset to enhance the dataset, which contains an average of multiple small target objects in the remote sensing images in the dataset; S3. The preprocessed data is imported into the improved YOLOv3 network for training.

2. The method for detecting small targets in remote sensing images based on improved YOLO deep learning according to claim 1, characterized in that, In step S2, the preprocessing operations on the remote sensing images in the dataset include: random scale transformation, random shifting, random left and right flipping Mosaic, and color dithering algorithm preprocessing.

3. The method for detecting small targets in remote sensing images based on improved YOLO deep learning according to claim 1, characterized in that, In the Ghost module, m original feature maps are generated using a single convolution. The operation formula Y′=X*f′ is used to optimize the f′ convolution kernel; where The parameters of the convolution kernel f′ are the same as those in a regular convolution kernel. Then, the following formula is used to perform a less computationally intensive convolution on the features in Y′, generating s feature maps that are the result of convolving the features generated in the previous step:

4. The method for detecting small targets in remote sensing images based on improved YOLO deep learning according to claim 1, characterized in that, In step S3, for the improved YOLOv3 network, the following formula is used for node feature fusion: In the formula, Conv represents the convolution operation; Re represents upsampling or downsampling of the feature map; Indicates that from node P 1,1 To node P 2,2 The weights of network connections are ensured to be greater than 0 using the ReLU function; ∈ is set to 10. -4 This is to avoid numerical instability by adding a number close to 0.

5. The method for detecting small targets in remote sensing images based on improved YOLO deep learning according to claim 1, characterized in that, The method for detecting small targets in remote sensing images also includes the following steps: A DA layer is introduced into the LKGCSP layer to generate the DALKGCSP layer. The DALKGCSP layer consists of a first 1*1 conv layer, a DA layer, n LKGB layers, a concat layer, and a second 1*1 conv layer connected in sequence. The input of the first conv layer is also connected to one of the inputs of the concat layer through a third 1*1 conv layer. The value of n is the same as the number of blocks in the corresponding DALKGCSP layer. The DA layer consists of a channel attention module and a spatial attention module connected in sequence. Each LKGB layer consists of two 7*7 convolutional layers and one 1*1 convolutional layer connected in sequence. First, the low-dimensional features of the input image are mapped to a high-dimensional space through the two 7*7 convolutional layers for convolution extraction, and then the 1*1 convolutional layer is used to reduce the dimensionality and output the result.

6. The method for detecting small targets in remote sensing images based on improved YOLO deep learning according to claim 5, characterized in that, The method for detecting small targets in remote sensing images also includes the following steps: The eighth and ninth LKGCSP layers are introduced into the FPN module to generate the BiFPN module. The output of the third LKGCSP layer is connected to one of the inputs of the first linear fusion layer. The output of the seventh LKGCSP layer is directly input to the detector on one hand, and connected to the other input of the first linear fusion layer through the conv layer on the other hand. The first linear fusion layer performs linear fusion on the input data and then inputs it to the eighth LKGCSP layer, which in turn inputs it to the detector. The output of the fourth LKGCSP layer is connected to one of the inputs of the second linear fusion layer. The output of the eighth LKGCSP layer is connected to the other input of the second linear fusion layer through the conv layer. The second linear fusion layer performs linear fusion on the input data and then inputs it to the ninth LKGCSP layer, which in turn inputs it to the detector.

Images