Remote sensing image target detection method fusing multi-scale context features and channel enhancement

Abstract

The application discloses a remote sensing image target detection method fusing multi-scale context features and channel enhancement, and the steps include: 1, pre-processing remote sensing image dataset; 2, constructing a neural network model based on cascade recursive convolution and attention mechanism; 3, using the model of step 2, constructing a multi-scale context feature enhancement network; 4, using sub-pixel convolution and adaptive sampling factor to design a spatial pyramid channel enhancement network to perform multi-scale feature fusion and detection. The application can solve the problem of various target scales in remote sensing target detection, improve the detection effect of multi-scale targets, and complete efficient detection of unique features of remote sensing image targets in complex backgrounds.

Description

Technical Field

[0001] This invention relates to remote sensing image target detection technology, specifically to a remote sensing image target detection method that integrates multi-scale contextual features and channel enhancement. Background Technology

[0002] Remote sensing (RS), literally meaning "perception from a distance," is a non-contact, long-range detection technology. In a broad sense, remote sensing refers to the art, science, and technology of recording, measuring, and interpreting images and digital pictures obtained through non-contact sensors (mainly spaceborne or airborne) to acquire reliable information about natural objects and the environment. With the continuous advancement of science and technology in recent years, remote sensing technology has developed rapidly. High-resolution satellites and airborne sensors have provided massive amounts of high-quality, high-resolution remote sensing images. Remote sensing images contain rich information about ground features, from which a wealth of valuable information can be extracted for scientific and technological research. In the field of digital image processing, remote sensing images, due to their unique imaging perspective, are widely used in areas such as building identification, resource exploration, crop pest and disease detection, urban planning, road defect detection, and military reconnaissance, making them an important spatial information resource.

[0003] Remote sensing image target detection is the identification and localization of specific categories of objects within an image. Research in this field focuses on improving the performance of detection tasks to achieve more accurate localization and more detailed classification. The primary condition for localization and recognition in remote sensing image target detection is feature extraction. Traditional feature extraction methods mainly employ feature extraction operators to obtain a series of basic geometric features such as edges, textures, and colors. Commonly used feature extraction operators include the Harris corner detection operator, Histogram of Oriented Gradients (HOG), Haar features, and Scale Invariant Feature Transform (SIFT). Based on this, machine learning methods such as template matching or Adaboost are used to achieve subsequent target detection tasks. However, remote sensing images have high resolution and large data volumes, making traditional detection methods relatively inefficient and resulting in high time and labor costs.

[0004] Deep learning is an artificial intelligence module that mimics how the human brain processes data and creates patterns for decision-making. It's a subset of machine learning in artificial intelligence (AI) that features networks capable of learning unsupervised from unstructured or unlabeled data. Over the past two decades, deep learning has seen tremendous development in both theory and application, with numerous major breakthroughs. It has been successfully applied to many computer application fields, including artificial intelligence, pattern recognition, data mining, natural language processing, speech recognition, image recognition, information retrieval, and bioinformatics, becoming a core technology in these areas. Compared to traditional algorithms, convolutional neural networks in deep learning can learn higher-level semantic information, thus exhibiting greater robustness and better suited for remote sensing image target detection tasks. Remote sensing images are characterized by high resolution, diverse scales, complex backgrounds, numerous small targets, and arbitrary target orientations. Fully exploring the immense information value contained within remote sensing images presents significant research opportunities for target detection technology.

[0005] Due to the complex backgrounds of remote sensing images, directly fusing features after feature extraction via a backbone network can easily lead to background interference at different scales, affecting subsequent detection performance. Current backbone network feature enhancement methods often add new modules after deep feature extraction to obtain a larger receptive field for the feature map, neglecting the crucial role of contextual information contained in shallow features in subsequent feature fusion. Traditional FPN methods also suffer from the loss of some channel information during feature fusion because the high channel count of deep features significantly decreases after 1x1 convolution. Summary of the Invention

[0006] This invention aims to address the shortcomings of existing technologies by proposing a remote sensing image target detection method that integrates multi-scale contextual features and channel enhancement. This method is intended to solve the problem of diverse target scale variations in remote sensing target detection, improve the detection effect of multi-scale targets, and achieve efficient detection of unique features of remote sensing images under complex backgrounds.

[0007] The present invention adopts the following technical solution to solve the technical problem:

[0008] The remote sensing image target detection method of the present invention, which integrates multi-scale contextual features and channel enhancement, is characterized by the following steps:

[0009] Step 1: Obtain the remote sensing image dataset with target bounding boxes, denoted as T = {T n |n=1,2,…,N}, where T n This represents the nth remote sensing image; N represents the number of remote sensing images.

[0010] Step 2: Construct a remote sensing image target detection network, including: feature extraction module f, multi-scale context enhancement module g, spatial pyramid channel enhancement module e, and detection head module u, and input the remote sensing image dataset T into the remote sensing image target detection network;

[0011] Step 2.1: The feature extraction network f includes: one directly connected block d and I skip connected blocks D, and a max pooling layer P is set between the directly connected block and the first skip connected block. max The direct connection block d sequentially includes a convolutional layer, a normalization layer, and an activation layer. Any skip connection block D is composed of multiple cascaded direct connection blocks and connected using a local residual method.

[0012] The input terminal of the feature extraction network f is the input terminal of the directly connected block d; the output terminal of the directly connected block d is connected to the max pooling layer P. max The input connection is used for the maximum pooling layer P. max The output terminal is connected to the input terminal of the first jump block D1, and a residual connection structure is adopted, that is: max pooling layer P max The output terminal is connected to the input terminal of the first jump connection block D1, and after passing through multiple direct connection blocks in the first jump connection block D1, it is connected to the input terminal of the second jump connection block D2. Simultaneously, the max-pooling layer P... max The output terminal skips the first jumper block D1 and is directly connected to the input terminal of the second jumper block D2;

[0013] The i-th jump block D i The output terminal is connected to the (i+1)th jumper block D i+1 The input terminal is connected, and it passes through the (i+1)th jumper block D. i+1 After multiple direct-connect blocks in the middle, it connects with the (i+2)th jump-connect block D. i+2 The input terminal is connected, and at the same time, the i-th jump block D i The output skips the (i+1)th jump block D i+1 Directly connected to the (i+2)th jump block D i+2 The input terminal is connected;

[0014] The nth remote sensing image T n After processing by the feature extraction network f, the I hop blocks D of the direct block d output feature maps F of different scales respectively. n ={F n,i |i=1,2,…,I,I+1}, where F n,1 F represents the feature map output by the directly connected block d. n,i Let i = 2, 3, ..., I, representing the (i-1)th jump block D. i Output feature map;

[0015] Step 2.2: The multi-scale context enhancement network g is used to acquire the nth remote sensing image T. n Contextual features;

[0016] Step 2.2.1: Transfer the i-th feature map F n,i The input is fed into the multi-scale context enhancement network g, and m+1 sets of features are first obtained through segmentation. in, Represents the i-th feature map F n,i The 0th feature, C0 is The number of channels, For the i-th feature map F n,i The (j+1)th feature, C j for The number of channels, and 0≤j≤m, H is the length of the feature map, W is the width of the feature map, and C is the input feature map F. n,i The number of channels;

[0017] Step 2.2.2: The multi-scale context enhancement network g uses equation (1) to construct cascaded recurrent convolution pairs for features. and The i-th feature map F is obtained through processing. n,i The (j+1)th order feature

[0018]

[0019] In equation (1), Dw represents the depthwise separable convolution operation, and ⊙ represents the matrix multiplication operation. Matrix addition operation, Ch is the channel transformation operation; Represents the i-th feature map F n,i The j-th order feature;

[0020] Step 2.2.3: The multi-scale context enhancement network g uses equation (2) to obtain the i-th feature map F. n,i Multiscale enhancement feature G n,i Thus, the (I+1)th feature map F is obtained. n,I+1 Multiscale enhancement feature G n,I+1 ;

[0021]

[0022] In equation (2), CA represents the channel attention mechanism. For the i-th feature map F n,i The m-th order feature;

[0023] Step 2.3: The spatial pyramid channel enhancement network e utilizes sub-pixel convolution and an adaptive sampling factor to enhance the multi-scale feature G. n,i Enhance channels and achieve efficient integration;

[0024] Step 2.3.1, Spatial pyramid channel enhancement network e to G n,i After performing subpixel convolution processing, the scale is magnified by a factor of z, and the number of channels is reduced by 1 / z. 2 The feature map is then transformed by channel transformation and compared with the (i-1)th feature map F. n,i-1 Multiscale enhancement feature G n,i-1 After fusion, a new multi-scale enhancement feature G′ is obtained. n,i-1 ;

[0025] Step 2.3.2: Construct constraints for α and β using equation (3), and set G... n,i Perform α-fold upsampling, and then... n,i-1 Perform a β-fold downsampling, and then fuse the results of the two methods to obtain G′. n,i-1 and G′ n,i intermediate layer features

[0026]

[0027] In equation (3), α is the upsampling factor and β is the downsampling factor. Obtain the feature map length and width after rounding down;

[0028] Step 2.3.3, the intermediate layer features After performing α-fold upsampling, and with G′ n,i-1 The features P of the (i-1)th remote sensing target detection network are obtained by fusion. n,i-1 ;

[0029] Step 2.3.4: Following the process in steps 2.3.1-2.3.2, process the (i+1)th feature map F. n,I+1 Multiscale enhancement feature G n,I+1 After convolution processing, it is combined with the features of the I-th intermediate layer. The results of β-fold downsampling are fused to obtain the (I+1)th feature map F. n,I+1 The remote sensing target detection network outputs features P n,I+1 ;

[0030] Step 3: Construct the detection head module u, which includes two branch predictors based on convolution operations with shared weights;

[0031] Step 3.1, for G n,I+1 After performing the convolution transformation, the output feature P of the (i+2)th remote sensing target detection network is obtained.n,I+2 , for P n,I+2 After channel adaptive downsampling, the output feature P of the (I+3)th remote sensing target detection network is obtained. n,I+3 ;

[0032] Step 3.2: The two branch predictors in the detection head module respectively predict P. n,i-1 ,P n,I+1 ,P n,I+2 ,P n,I+3 The process involves two branches: one predictor outputs the category of each target bounding box in each remote sensing image, and the other predictor outputs the location information of each target bounding box in each image.

[0033] Step 3.3: Based on the output of the detection head module and the target bounding box in the remote sensing image, calculate the Focal loss (classification loss) and smoothness. L1 The regression loss is used to backpropagate the remote sensing image target detection network using the gradient descent method to update the network parameters until the loss converges, thereby obtaining the optimal remote sensing image target detection model for classification and recognition of remote sensing images.

[0034] The present invention provides an electronic device, including a memory and a processor, wherein the memory is used to store a program that supports the processor in executing the remote sensing image target detection method, and the processor is configured to execute the program stored in the memory.

[0035] The present invention discloses a computer-readable storage medium on which a computer program is stored, wherein the computer program, when executed by a processor, performs the steps of the remote sensing image target detection method.

[0036] Compared with existing technologies, the beneficial effects of this invention are reflected in:

[0037] 1. Compared with previous remote sensing image target detection technologies, this invention focuses on analyzing the problems of complex backgrounds and diverse target scales in remote sensing images. It uses a detection method that integrates multi-scale contextual features and channel enhancement to improve detection accuracy.

[0038] 2. This invention uses cascaded recurrent convolution and attention mechanisms to construct a multi-scale context feature enhancement network. By adding a cascaded structure to the recurrent gated convolution, information before interaction is retained, thus obtaining richer context features. The channel number segmentation method is changed to reduce the number of channels, thereby reducing the required computation and network complexity. By adding a channel attention mechanism to obtain weight coefficients, important features after spatial interaction are enhanced, and redundant information is suppressed, thus obtaining richer and more accurate target feature information and improving the detection accuracy of multi-scale targets.

[0039] 3. This invention utilizes sub-pixel convolution and adaptive sampling factors to design a new spatial pyramid feature fusion method. Based on the output features of the multi-scale context enhancement module, sub-pixel convolution is used to reconstruct features, reducing the loss due to excessive channel changes and achieving channel enhancement effect. By using channel adaptive sampling factors, an intermediate layer is added between adjacent feature layers, thereby reducing the information loss caused by feature map scale changes during sampling and improving the target detection effect in complex backgrounds. Attached Figure Description

[0040] Figure 1 This is a diagram showing the overall structure of the detection network that integrates multi-scale contextual features and channel enhancement in this invention.

[0041] Figure 2 This is a diagram of the feature extraction network structure in this invention;

[0042] Figure 3 This is a diagram of the multi-scale context enhancement network structure in this invention;

[0043] Figure 4 This is a diagram of the spatial pyramid channel enhancement network structure of the present invention. Detailed Implementation

[0044] In this embodiment, a remote sensing image target detection method integrating multi-scale contextual features and channel enhancement is proposed to address the problems of complex backgrounds and diverse target scales in remote sensing images. The algorithm network mainly consists of a multi-scale contextual enhancement module and a feature pyramid channel enhancement module. The former can adaptively enhance the spatial features extracted from convolutional blocks of different scales in the backbone network, obtaining more contextual information and reducing the impact of background information on subsequent spatial pyramid fusion. The latter aims to enhance the channel information in multi-scale spatial pyramid fusion, mitigating information loss due to feature map channel dimensionality changes and upsampling / downsampling. Specifically, the method proceeds according to the following steps:

[0045] Step 1: Obtain the remote sensing image dataset with target bounding boxes, denoted as T = {T n |n=1,2,…,N}, where T n This represents the nth remote sensing image; N represents the number of remote sensing images, and the remote sensing image dataset is preprocessed.

[0046] In this embodiment, the input image is first randomly flipped with a 50% probability, then the image is uniformly reset to 800*800 size using linear interpolation, then the image is padded to a multiple of 32, and finally the image is uniformly normalized to facilitate convergence.

[0047] Step 2: Construct a remote sensing image target detection network, including: feature extraction module f, multi-scale context enhancement module g, spatial pyramid channel enhancement module e, and detection head module u. Input the remote sensing image dataset T into the remote sensing image target detection network. The overall structure of the remote sensing image target detection network is as follows: Figure 1 As shown;

[0048] The general workflow is as follows: First, obtain the feature map F extracted by the backbone network. n,i This is used as input to obtain more information through a multi-scale contextual feature enhancement module, resulting in multi-scale enhanced features G. n,i Then, it undergoes 1*1 convolution (Conv) for channel reduction, and finally, it is fused using the feature pyramid channel enhancement module to obtain P. n,i-1 and P n,I+1 In G n,I+1 Based on this, P is obtained through 3x3 convolution. n,I+2 P n,I+2 P was obtained after 0.75x downsampling. n,I+3 Finally, for feature map P n,i-1 ,P n,I+1 ,P n,I+2 ,P n,I+3 Classify and locate.

[0049] Step 2.1, Feature extraction network f as follows Figure 2 As shown, it includes: one direct connection block d and I skip connection blocks D, and there is a max pooling layer P between the direct connection block and the first skip connection block. max The directly connected block d contains, in sequence, a convolutional layer, a normalization layer, and an activation layer. Any skip block D is composed of multiple directly connected blocks cascaded together and connected using a local residual method.

[0050] The input of the feature extraction network f is the input of the directly connected block d; the output of the directly connected block d is connected to the max pooling layer P. max The input connection is used for the maximum pooling layer P. max The output terminal is connected to the input terminal of the first jumper block D1, and a residual connection structure is adopted, that is: max pooling layer P max The output terminal is connected to the input terminal of the first jump connection block D1, and after passing through multiple direct connection blocks in the first jump connection block D1, it is connected to the input terminal of the second jump connection block D2. Simultaneously, the max-pooling layer P... max The output terminal skips the first jumper block D1 and is directly connected to the input terminal of the second jumper block D2;

[0051] The i-th jump block D i The output terminal is connected to the (i+1)th jumper block D i+1 The input terminal is connected, and it passes through the (i+1)th jumper block D. i+1After multiple direct-connect blocks in the middle, it connects with the (i+2)th jump-connect block D. i+2 The input terminal is connected, and at the same time, the i-th jump block D i The output skips the (i+1)th jump block D i+1 Directly connected to the (i+2)th jump block D i+2 The input terminal is connected;

[0052] The nth remote sensing image T n After processing by the feature extraction network f, the I hop blocks D of the directly connected block d output feature maps F of different scales respectively. n ={F n,i |i=1,2,…,I,I+1}, where F n,1 F represents the feature map output by the directly connected block d. n,i Let i = 2, 3, ..., I, representing the (i-1)th jump block D. i Output feature map;

[0053] In this example, I is set to 4, and F is selected. n,5 ,F n,4 ,F n,3 The output feature map, proposed as a feature, is based on an input image of size 800*800, with 3 channels of RGB. The first layer of direct-connected blocks has 64 output channels, a kernel size of 7, and a stride of 2. Each skip-connected block in the middle four layers consists of three direct-connected blocks, with output channels of (64, 64, 256), (128, 128, 512), (256, 256, 1024), and (512, 512, 2048), respectively, and repetitions of 3, 4, 6, and 3 times per layer. One convolutional kernel in the middle of each of the three direct-connected blocks has a size of 3, while the remaining kernels have a size of 1, reducing the number of channels and computational cost. The three repetitions of the first layer skip-connected block are identical, and all convolutional operations have a stride of 1. Starting from the second layer skip-connected block, the first repetition of the 3*3 convolutional operation in each layer has a stride of 2, while the remaining convolutions have a stride of 1. The input image T is selected from the training dataset. n The feature maps F at different scales are obtained sequentially through the feature extraction network f. n,3 ,F n,4 ,F n,5 The number of channels is 512, 1024 and 2048 respectively. Then, the feature map is input into a multi-scale contextual feature enhancement network to perform spatial information interaction of the second, third and fourth order to enhance the feature extraction effect.

[0054] Step 2.2: Construct a multi-scale contextual enhancement network g, such as... Figure 3 As shown, this is used to acquire the nth remote sensing image T. n Contextual features;

[0055] Step 2.2.1: Transfer the i-th feature map F n,i The input is fed into a multi-scale context enhancement network g, where m+1 sets of features are first obtained through segmentation. in, Represents the i-th feature map F n,i The 0th feature, C0 is The number of channels, For the i-th feature map F n,i The (j+1)th feature, C j for The number of channels, and H is the length of the feature map, W is the width of the feature map, and C is the input feature map F. n,i The number of channels.

[0056] Step 2.2.2: The multi-scale context enhancement network g uses equation (1) to construct cascaded recurrent convolution pairs for features. and The i-th feature map F is obtained through processing. n,i The (j+1)th order feature

[0057]

[0058] In equation (1), Dw represents the depthwise separable convolution operation, and ⊙ represents the matrix multiplication operation. Matrix addition operation, Ch is the channel transformation operation; Represents the i-th feature map F n,i The j-th order feature.

[0059] Step 2.2.3: The multi-scale context enhancement network g uses equation (2) to obtain the i-th feature map F. n,i Multiscale features G n,i Thus, the (I+1)th feature map F is obtained. n,I+1 Multiscale features G n,I+1 ;

[0060]

[0061] In equation (2), For the i-th feature map F n,i The m-th feature, CA represents the channel attention mechanism, which uses parallel adaptive max pooling and average pooling, summed after a 16-fold channel scaling operation, and then obtains the channel attention weight coefficients through the Sigmoid activation function.

[0062] In this example, taking the 3rd-order interaction as an example, one branch obtains the C / 8 information after the feature layer is normalized, and the remaining 7C / 8 are separated by the depthwise separable convolution. Then, the spatial interaction with the feature map channels of C / 8, C / 4 and C / 2 is recursively executed 3 times, and finally fused into a result with the number of channels C.

[0063] Step 2.3: The spatial pyramid channel enhancement network e utilizes sub-pixel convolution and adaptive sampling factors to enhance multi-scale features G. n,i Channel enhancement and efficient fusion are performed, with a modular structure as follows: Figure 4 As shown;

[0064] Step 2.3.1: Spatial pyramid feature enhancement network e on G n,i After performing subpixel convolution processing, the scale is magnified by a factor of z, and the number of channels is reduced by 1 / z. 2 The feature map is then transformed by channel transformation and compared with the (i-1)th feature map F. n,i-1 Multiscale features G n,i-1 After fusion, a new multi-scale feature G is obtained. n ′ ,i-1 ;

[0065] In this example, i = 3, 4, 5, G n,5 The channel count is 2048, r is 2, after subpixel convolution the channel count becomes 512, and then after channel transformation it becomes 1024 and G n,4 To integrate.

[0066] Step 2.3.2: Construct constraints for α and β using equation (3), and set G... n,i Perform α-fold upsampling, and then... n,i-1 Perform a β-fold downsampling, and then fuse the results of the two methods to obtain G′. n,i-1 and G′ n,i intermediate layer features

[0067]

[0068] In equation (3), α is the upsampling factor and β is the downsampling factor. Obtain the feature map length and width after rounding down.

[0069] Step 2.3.3, Intermediate Layer Features After performing α-fold upsampling, and with G′ n,i-1 The features P of the (i-1)th remote sensing target detection network are obtained by fusion. n,i-1 .

[0070] Step 2.3.4: Following the process in steps 2.3.1-2.3.2, process the (i+1)th feature map F.n,I+1 Multiscale features G n,I+1 After convolution processing, it is combined with the features of the I-th intermediate layer. The results of β-fold downsampling are fused to obtain the (I+1)th feature map F. n,I+1 The remote sensing target detection network outputs features P n,I+1 ;

[0071] In this example, feature map G n,5 The scale size is 25*25, G n,4 The feature map size is 50*50, the adaptive upsampling factor is 1.42, the downsampling factor is 0.715, and the intermediate layer feature map size is 37.

[0072] Step 3: Construct the detection head module u, which includes two branch predictors based on convolution operations with shared weights;

[0073] Step 3.1, for G n,I+1 After performing the convolution transformation, the output feature P of the (i+2)th remote sensing target detection network is obtained. n,I+2 , for P n,I+2 After channel adaptive downsampling, the output feature P of the (I+3)th remote sensing target detection network is obtained. n,I+3 .

[0074] Step 3.2: The two branch predictors in the detection head module respectively target P. n,i-1 ,P n,I+1 ,P n,I+2 ,P n,I+3 The process involves two branches: one predictor outputs the category of each target bounding box in each remote sensing image, and the other predictor outputs the location information of each target bounding box in each image.

[0075] In this example, each feature layer uses 3 scales (2^35). 0 ,2 1 / 3 ,2 2 / 3 The dataset contains 9 target prediction boxes and 3 ratios (1:2, 1:1, 2:1). The kernel size is 3 and the stride is 1. The number of channels in the pre-convolutional feature map is 256. The number of channels in the final convolutional feature map of the first branch classification is W×H×180, and the number of channels in the second branch regression is W×H×36.

[0076] Step 3.3: Based on the output of the detection head module and the target bounding box in the remote sensing image, calculate the Focal loss (classification loss) and smoothness loss. L1The regression loss is used to backpropagate the remote sensing image target detection network using the gradient descent method to update the network parameters until the loss converges, thereby obtaining the optimal remote sensing image target detection model, which is used to classify and recognize remote sensing images.

[0077] In this example, the final input to the detection head is P. n,3 ,P n,4 ,P n,5 ,P n,6 ,P n,7 ;P n,6 For F n,5 The feature map after performing a convolutional kernel size of 3 and a stride of 2 has its number of channels reduced from 2048 to 256. (P) n,7 For P n,6 The result after performing a 0.75x downsampling is scaled 96 times compared to the original feature map. When matching positive and negative samples, each predicted box is compared with the pre-labeled ground truth boxes. If the overlap IoU between the two boxes is greater than 0.5, it is a positive sample. If the IoU between a predicted box and all ground truth boxes is less than 0.4, it is a negative sample. The rest are discarded.

[0078] In this embodiment, an electronic device includes a memory and a processor. The memory stores a program that supports the processor in executing the above-described method, and the processor is configured to execute the program stored in the memory.

[0079] In this embodiment, a computer-readable storage medium stores a computer program, which is executed by a processor to perform the steps of the above method.

[0080] The network model was tested on the DIOR dataset, and the mAP was 67.8%, which is 2.1% higher than the benchmark network RetinaNet, as shown in Table 1.

[0081] Table 1. Experimental comparison on the DIOR dataset.

[0082]

[0083] As can be seen from the experimental data in Table 1, when the multi-scale context enhancement network was added to the baseline network, the mAP changed from 65.7% to 67.3%, and the detection accuracy improved by 1.6%. Therefore, the multi-scale context enhancement network alleviated the problem of diverse target scales in remote sensing images. At the same time, after introducing sub-pixel convolution and adaptive sampling factor, the accuracy improved by 0.8% compared with the baseline network, which enhanced the feature information of the target and improved the target detection effect in remote sensing images under complex backgrounds.

Claims

1. A remote sensing image target detection method fusing multi-scale context features and channel enhancement, characterized in that, is carried out as follows: Step 1, obtain a remote sensing image dataset with a target real box, denoted as T = {T n |n=1,2,…,N},wherein, T n represents the nth remote sensing image; N represents the number of remote sensing images; Step 2, constructing a remote sensing image target detection network, including: a feature extraction module f, a multi-scale context enhancement module g, a spatial pyramid channel enhancement module e, and a detection head module u, and inputting the remote sensing image data set T into the remote sensing image target detection network; Step 2.1, the feature extraction network f comprises one direct connection block d and I skip connection blocks D, and a max pooling layer P is arranged between the direct connection block and the first skip connection block max The direct connection block d comprises, in sequence, a convolution layer, a normalization layer and an activation layer, any one skip connection block D is formed by cascading multiple direct connection blocks, and is connected in a local residual manner; The input end of the feature extraction network f is the input end of the direct connection block d; the output end of the direct connection block d is connected with the input end of the max pooling layer P max , the output end of the max pooling layer P max is connected with the input end of the first skip connection block D1, and a residual connection structure is adopted, that is, the output end of the max pooling layer P max is connected with the input end of the first skip connection block D1 and is connected with the input end of the second skip connection block D2 after passing through a plurality of direct connection blocks in the first skip connection block D1, and meanwhile, the output end of the max pooling layer P max is directly connected with the input end of the second skip connection block D2 by skipping the first skip connection block D1. The output end of the i th jump block D i is connected with the input end of the i+1 th jump block D i+1 , and after passing through a plurality of direct connection blocks in the i+1 th jump block D i+1 , is connected with the input end of the i+2 th jump block D i+2 , while the output end of the i th jump block D i directly connects with the input end of the i+2 th jump block D i+2 , skipping the i+1 th jump block D i+1 . The nth remote sensing image T n After processing by the feature extraction network f, the I jump connection blocks D of the direct connection block d respectively output feature maps F of different scales n ={F n,i |i=1,2,…,I,I+1}, where F n,1 represents the feature map output by the direct connection block d, F n,i , i=2,3,…I, represents the feature map output by the i-1th jump connection block D i ​ Step 2.2, the multi-scale context enhancement network g, is used to obtain the context features of the nth remote sensing image T n . Step 2.2.1, the i-th feature map F n,i is input into the multi-scale context enhancement network g, and m+1 groups of features are obtained by cutting processing wherein, represents the 0-th group of features of the i-th feature map F n,i , C0is the number of channels of F , C n,i is the j+1-th group of features of the i-th feature map F j , C is the number of channels of F , H is the length of the feature map, W is the width of the feature map, and C is the number of channels of the input feature map F n,i . Step 2.2.2, the multi-scale context enhancement network g constructs a cascaded recursive convolution pair of features using formula (1) and is processed to obtain the (j+1)th feature of the ith feature map F n,i ​ In formula (1), Dw is a depth separable convolution operation, is a matrix multiplication operation, a matrix addition operation, Ch is a channel transform operation. represents the jth-order feature of the ith feature map F n,i . Step 2.2.3, the multi-scale context enhancement network g obtains the i-th feature map F n,i of the multi-scale enhanced feature G n,i by using formula (2), thereby obtaining the i+1-th feature map F n,I+1 of the multi-scale enhanced feature G n,I+1 ; In formula (2), CA represents a channel attention mechanism, is the mth-order feature of the ith feature map F n,i is the mth-order feature of the ith feature map F Step 2.3, the spatial pyramid channel enhancement network e utilizes sub-pixel convolution and adaptive sampling factors on the multi-scale enhanced features G n,i Channel enhancement and efficient fusion are performed; Step 2.3.1, spatial pyramid channel enhancement network e enhances G n,i After sub-pixel convolution processing, a new multi-scale enhanced feature G' is obtained 2 After sub-pixel convolution processing, a new multi-scale enhanced feature G' is obtained n,i-1 After sub-pixel convolution processing, a new multi-scale enhanced feature G' is obtained n,i-1 After sub-pixel convolution processing, a new multi-scale enhanced feature G' is obtained n,i-1 ; Step 2.3.2, construct constraints for a and β using formula (3), G n,i Perform a times up-sampling, G n ,i-1 Perform β times down-sampling, then fuse the results of both to obtain G' n,i-1 n,i Intermediate layer features of G and G' ​​ In formula (3), a is an up-sampling factor, and β is a down-sampling factor. obtaining a value obtained by rounding down a length and a width of the feature map; Step 2.3.3, the intermediate layer feature After α times up-sampling, G' n,i-1 Fusion is performed to obtain the i-1th remote sensing target detection network output feature P n,i-1 ; Step 2.3.4, according to the process of steps 2.3.1-2.3.2, the (I+1)th feature map F n,I+1 of the multi-scale enhanced feature G n,I+1 of the (I+1)th feature map F after the convolution processing is fused with the result of the β times down-sampling of the Ith intermediate layer feature n,I+1 , thereby obtaining the remote sensing target detection network output feature P n,I+1 of the (I+1)th feature map F Step 3, constructing a detection head module u, including two branch predictors based on convolution operation with weight sharing; Step 3.1, to G n,I+1 After convolution transformation, the (I+2)th remote sensing target detection network output feature P is obtained n,I+2 , P n,I+2 After channel adaptive down-sampling, the (I+3)th remote sensing target detection network output feature P n,I+3 is obtained; Step 3.2, two branch predictors in the detection head module respectively process P n,i-1 ,P n,I+1 ,P n,I+2 ,P n,I+3 , and one branch predictor outputs the category of each target prediction frame in each remote sensing image, and the other branch predictor outputs the position information of each target prediction frame in each image; Step 3.3, according to the output result of the detection head module and the target real box of the remote sensing image, calculate the Focal loss classification loss and smooth L1 regression loss, so as to use gradient descent method to update the network parameters of the remote sensing image target detection network by back propagation until the loss converges to the position, so as to obtain the optimal remote sensing image target detection model for realizing the classification and recognition of remote sensing image.

2. An electronic device comprising a memory and a processor, characterized in that The memory is used to store a program supporting the processor to execute the remote sensing image target detection method of claim 1, and the processor is configured to execute the program stored in the memory.

3. A computer-readable storage medium having stored thereon a computer program, characterized in that The computer program is executed by the processor to perform the steps of the remote sensing image target detection method of claim 1.

Images