Remote sensing image target detection method based on MCD-YOLO network

Abstract

The invention discloses a remote sensing image target detection method based on MCD-YOLO network, comprising the following steps: step 1, obtaining a public remote sensing data set and preprocessing; step 2, configuring a training environment; step 3, constructing an MCD-YOLO network model, wherein the MCD-YOLO network model uses a YOLOv8 network as a basic network, replaces the C2f module in the backbone network of the YOLOv8 network with a C2f_LVMB module as a feature extraction module, and incorporates a CA attention mechanism into the SPPF module; step 4, training and verifying the MCD-YOLO network model using the preprocessed data set; step 5, performing target detection by using a pre-trained MCD-YOLO network model using a remote sensing image to be detected as input. This method not only effectively reduces the amount of calculation and improves detection efficiency, but also effectively improves model performance by extracting information of different scales.

Description

Technical Field

[0001] The present invention relates to the field of computer vision technology, and in particular to remote sensing image target detection based on an MCD-YOLO network. Background Art

[0002] With the rapid development of remote sensing technology, the spatial resolution, spectral resolution and temporal resolution of earth observation satellites have been continuously improved. Remote sensing technology has now become the main technical means to collect information on the earth's surface and its dynamic changes. Through remote sensing technology, global resource and environmental information can be obtained in a large range, in all directions and at a high rate. At the same time, it also plays an increasingly important role in military detection, civilian life and natural disaster relief. Therefore, the information acquisition in remote sensing images is particularly important.

[0003] In recent years, deep learning technology has been widely used in the field of remote sensing. It can autonomously and efficiently learn complex features in remote sensing images and quickly recognize images. Convolutional neural networks have the characteristics of weight sharing and translation and scale invariance, and have become one of the main deep learning technologies in the field of remote sensing image target detection. In addition to convolutional neural networks, networks such as GAN (Generative Adversarial Network), RNN (Recurrent Neural Network) and AE (Auto Encoder) are also widely used in the field of target detection. Deep learning technology has brought a lot of convenience to the field of target detection.

[0004] However, there are still many challenges in remote sensing images. First, the objects are oriented in different directions, and traditional detection algorithms lack robustness to feature differences caused by different orientations. Second, the background is complex and the context information is limited, so traditional algorithms are prone to misdetecting objects with similar features. Finally, the image resolution is large and the scales of targets in the image are different. In addition, traditional convolutional neural networks consume a lot of computing resources, which makes it difficult to deploy and smoothly implement detection tasks in hardware environments with limited computing power. Therefore, how to make the model lightweight is also a problem that must be solved. Summary of the invention

[0005] In view of the shortcomings of the prior art, the present invention proposes a remote sensing image target detection method based on the MCD-YOLO network, which effectively improves the model performance by extracting information of different scales.

[0006] In order to solve the above technical problems, the technical solution of the present invention is:

[0007] A remote sensing image target detection method based on MCD-YOLO network includes the following steps:

[0008] Step 1: Obtain and preprocess the public remote sensing dataset;

[0009] Step 2: Configure the training environment;

[0010] Step 3, construct an MCD-YOLO network model, wherein the MCD-YOLO network model uses the YOLOv8 network as the basic network, replaces the C2f module in the backbone network of the YOLOv8 network with the C2f_LVMB module as the feature extraction module, and integrates the CA attention mechanism into the SPPF module in the backbone network; replaces the original detection head with a lightweight rotating box detection head in the head network of the YOLOv8 network;

[0011] Step 4: Use the preprocessed data set to train and verify the MCD-YOLO network model;

[0012] Step 5: Use the remote sensing image to be detected as input to perform target detection through the pre-trained MCD-YOLO network model.

[0013] Preferably, the preprocessing method is: processing the high-resolution image by segmentation, thereby dividing the image into blocks of fixed size.

[0014] In the above technical solution, the dataset is segmented into smaller sizes and then input into the model for training. Processing high-resolution images by segmentation can effectively improve the phenomenon of missed detection of small objects.

[0015] Preferably, in step 3, four C2f modules are provided in the backbone network of the original YOLOv8 network, and the last two are replaced by C2f_LVMB modules.

[0016] Preferably, the C2f_LVMB module uses a VSSBlock module to replace the Bottleneck module in the C2f module.

[0017] In the above technical solution, the VSSBlock module of VMamba is introduced and integrated with C2f to form the C2f_LVMB module. C2f_LVMB improves the model's feature extraction capability, enabling it to better capture the complex features of input data and improve the model's ability to recognize complex patterns.

[0018] Preferably, the method for extracting features by the C2f_LVMB module is:

[0019] First, use 1*1 convolution to change the number of channels i of the input feature map to the number of output channels m specified in the model configuration file, then use the Split function to divide the input into two branches with a channel number of 0.5*m. One branch extracts image features through n VSSBlock modules, and then connects with the other branch through n residual connections to obtain a feature map with a channel number of (n+2)*0.5m. Finally, use 1*1 convolution to change the number of channels to m to obtain the output.

[0020] Preferably, the method for constructing the lightweight rotating box detection head is: using group convolution to replace the standard convolution in the original detection head to obtain a lightweight detection head, and then adding an angle branch with an output channel number of 1 on the basis of the lightweight detection head, and combining the sigmoid activation function to map the output to the required angle range, and finally forming the output together with the regression information and category confidence, thereby constructing a lightweight rotating box detection head.

[0021] Preferably, the number of groups g of the grouped convolution is set to one eighth of the number of input channels.

[0022] Preferably, the grouped convolution divides the input data with a channel number of c1 into g groups for separate convolutions, where the number of channels in each group is one-g of the input channel number, i.e., c1 / g, and the depth of each group of convolution kernels is also changed to one-g of the input channel number, i.e., c1 / g, and the number of convolution kernels in each group is changed to one-g of the output channel number c2. The convolution kernels of each group are then convolved with the input data in their corresponding groups to obtain the output data, which are then combined in a splicing manner, and the final output channel number is still c2.

[0023] In the above technical solution, the input data is: H×W×c1, and the convolution kernel size is K×K, the channel is c1, and there are c2 convolution kernels in total. Assuming that the output resolution remains unchanged, g identical convolution processes are operated in parallel through grouped convolution. The input data of each process is H×W×c1 / g, the convolution kernel size is K×K×c1 / g, there are c2 / g in total, and the output data is H×W×c2 / g. The amount of calculation and parameter of grouped convolution is 1 / g of that of standard convolution.

[0024] Preferably, the MCD-YOLO network uses MPDIoU as the loss function, and the algorithm is as follows:

[0025] Define two arbitrary convex traits: , the width and height of the input image: ;

[0026] For A and B, Represents the coordinates of the upper left corner and lower right corner of A;

[0027] Represents the coordinates of the upper left corner and lower right corner of B;

[0028]

[0029]

[0030] .

[0031] In the above technical solution, in order to improve the detection effect of the lightweight model, the present invention adopts a new IOU calculation method MPDIoU. It is a new bounding box similarity metric based on the minimum point distance, which includes all relevant factors considered in the existing loss function, namely, overlapping or non-overlapping areas, center point distance, and width and height deviations, while simplifying the calculation process.

[0032] Preferably, the training method of the MCD-YOLO network is: dividing the data set into a training set and a validation set according to a ratio of 7:3, setting the training coefficient to 400 rounds, inputting 8 pictures for each training, observing the training process in real time through wandb during the training process, and saving the trained weights after the training is completed.

[0033] The present invention has the following characteristics and beneficial effects:

[0034] (1) In order to solve the problem of diverse orientations and multi-scales of objects in remote sensing images, the invention proposes to merge the VSSBlock module of VMamba with the C2f module to form the C2f_LVMB module, and introduce it into the backbone. The C2f-LVMB module obtains information in multiple directions by scanning the input in four directions, and can extract information at different scales.

[0035] (2) In order to improve the detection accuracy of the model, the present invention introduces the CA attention mechanism in SPPF. The CA attention mechanism effectively encodes spatial information into channel attention through global pooling, direction encoding and attention fusion, thereby enhancing the feature expression ability and improving the model performance. Without significantly increasing the computational complexity, the CA attention mechanism can better capture and utilize important information in the input feature map.

[0036] (3) In order to reduce the amount of model calculation, the present invention designs a lightweight detection head and constructs a lightweight rotating frame detection head for rotating targets. In addition, in order to improve the detection effect after lightweighting, the invention introduces a new iou calculation method MPDIoU, which is more effective in processing the shape and position errors of the detection frame. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0038] Figure 1 A model structure diagram of remote sensing image target detection based on the MCD-YOLO network provided in this embodiment.

[0039] Figure 2 for Figure 1 C2f_LVMB module structure diagram.

[0040] Figure 3 for Figure 1 The structure diagram of the SPPF_CA module.

[0041] Figure 4 for Figure 1 Structural diagram of the lightweight rotating frame detection head.

[0042] Figure 5 This is a comparison chart of the detection effects of the first case in this embodiment.

[0043] Figure 6 This is a comparison chart of the detection effects of the second case in this embodiment.

[0044] Figure 7 This is a comparison chart of the detection effects of the third case in this embodiment. DETAILED DESCRIPTION

[0045] It should be noted that, in the absence of conflict, the embodiments of the present invention and the features in the embodiments may be combined with each other.

[0046] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below with reference to the accompanying drawings.

[0047] On the contrary, the present invention covers any substitution, modification, equivalent method and scheme made on the essence and scope of the present invention as defined by the claims. Further, in order to make the public have a better understanding of the present invention, some specific details are described in detail in the detailed description of the present invention below. Those skilled in the art can fully understand the present invention without the description of these details.

[0048] The present invention provides a remote sensing image target detection method based on an MCD-YOLO network, comprising the following steps:

[0049] Step 1: Obtain and preprocess the public remote sensing dataset.

[0050] Specifically, the first step is to obtain the remote sensing data set. This embodiment uses the DOTAv1.5 data set. This data set is an improvement and extension of the DOTAv1.0 data set. It uses the same images as DOTAv1.0, but also annotates extremely small targets (less than 10 pixels). In addition, a new category "container crane" is added. The images come from different sensors and platforms, including satellite images and drone images, with diverse imaging conditions and scenes. The data set contains a total of 2,806 remote sensing images, including 16 types of targets, including aircraft, ships, storage tanks, baseball fields, tennis courts, basketball courts, athletic fields, ports, bridges, large vehicles, small vehicles, helicopters, roundabouts, football fields, swimming pools, container cranes, a total of 403,318 instances.

[0051] It is understandable that the DOTA dataset is used to train targets in aerial images, where the highest resolution of the images is up to 20,000 x 20,000 pixels. Due to computational limitations and the small and dense nature of the objects in these images, direct training is challenging. To solve this problem, the dataset needs to be segmented into smaller sizes and then input into the model for training. Processing high-resolution images by segmentation can effectively improve the phenomenon of missed detection of small targets. The segmented image usually uses a sliding window method to divide the image into blocks of fixed size. In this embodiment, the image is divided into a fixed size of 1024*1024 pixels, and adjacent blocks overlap by 200 pixels to maintain the integrity of objects across block boundaries.

[0052] Step 2: Configure the training environment.

[0053] Specifically, create the DOTA.yaml file of the data set, then write the paths of train and val and all the category information in this data set under the DOTA.yaml file, and then reasonably set the size of the input image, the number of trainings, and the category of the detected object according to the computer's memory and graphics card size. In addition, it is also necessary to consider whether the built environment is compatible with the graphics card. The configuration environment in this embodiment is: cuda11.8, deep learning framework pytorch1.12.1, Intel core i9-13900ks CPU, 128G memory, GPU is NVIDIA GeForce RTX 4090, and the video memory is 24G.

[0054] Step 3: Build the MCD-YOLO network model.

[0055] Specifically, Figure 1As shown in the figure, the MCD-YOLO network model uses the YOLOv8 network as the basic network and is divided into three parts: the backbone network, the neck and the head. The backbone network extracts features from the input through operations such as convolution and pooling. The neck fuses the feature maps extracted by the backbone network from bottom to top and from top to bottom. By fusing feature maps of different scales, the detailed information of low-level features and the semantic information of high-level features can be comprehensively utilized to improve the detection accuracy. Finally, the G-OBB detection head is used in the head to obtain the information of the detected object we need.

[0056] Further configuration of this embodiment is as follows: Figure 1 and Figure 2 As shown, the C2f module in the backbone network of the YOLOv8 network is replaced with a C2f_LVMB module as a feature extraction module. In this embodiment, the last two C2f modules of the original backbone network are replaced with C2f_LVMB modules.

[0057] Specifically, C2f_LVMB introduces the VSSBlock module of VMamba. Compared with the original C2f, C2f_LVMB uses VSSBlock instead of Bottleneck, which enables the model to extract features in different directions and reduces the amount of calculation.

[0058] Furthermore, C2f_LVMB improves the model's feature extraction capabilities, enabling it to better capture the complex features of input data and improve the model's ability to recognize complex patterns. The structure of the C2f_LVMB module is shown in the attached figure. Figure 2 As shown in (a), first, a 1*1 convolution is used to change the number of channels i of the input feature map to the number of output channels m specified in the model configuration file. Then, the Split function is used to divide the input into two branches with a channel number of 0.5*m. One branch extracts image features through n VSSBlock modules, and then is concatenated with the other branch through n residual connections to obtain a feature map with a channel number of (n+2)*0.5m. Finally, a 1*1 convolution is used to change the number of channels to m to obtain the output.

[0059] The VSSBlock module can integrate position information from different directions through four scanning methods. Its structure is shown in the attached figure. Figure 2In (b), the main structure of the module is a residual connection and SS2D (2D-Selective-Scan) module. SS2D processes data in three steps: cross scanning, selective scanning of the S6 module, and cross merging. Given the input data, SS2D first expands the input image blocks into sequences along four different traversal paths, then processes each sequence separately using S6 blocks in parallel, and finally reshapes and merges the resulting sequences to form an output graph. By adopting a complementary one-dimensional traversal path, SS2D enables each pixel in the image to effectively integrate information from all other pixels in different directions, thereby facilitating the establishment of a global receptive field in two-dimensional space.

[0060] Furthermore, the S6 module is a state-space model, which is used to convert the hidden state of the previous time step into With the current input Combined to get the current hidden state , and use Producing Output , which can be expressed mathematically as:

[0061]

[0062]

[0063] in

[0064]

[0065]

[0066]

[0067] In the formula, , , is a parameter matrix, using parameters After discretization, we get , , Three discretized parameter matrices. The selective scanning mechanism in the S6 module refers to the state transition All from input data This demonstrates that the state-space model is able to exploit contextual information in the input data, ensuring that the weights in the mechanism are always dynamic.

[0068] The present invention is further configured as follows: Figure 1 and Figure 3 As shown in Figure 2, the CA attention mechanism is integrated into the SPPF module in the backbone network. The structure of the SPPF_CA module is shown in the attached figure. Figure 3As shown in (a) of Figure 2. By introducing the CA attention mechanism, the model is able to model channel attention while capturing long-range dependencies with precise location information. The structural diagram of the CA attention mechanism is shown in the attached figure. Figure 3 As shown in (b).

[0069] Coordinate attention decomposes channel attention into two one-dimensional feature encoding processes that aggregate features along two spatial directions respectively. In this way, long-range dependencies can be captured along one spatial direction while retaining precise position information along the other spatial direction. The resulting feature maps are then encoded into a pair of direction-aware and position-sensitive attention maps, respectively, which can be complementarily applied to the input feature maps to enhance the representation of objects of interest.

[0070] VSSBlock can capture long-range dependencies and contextual information through dynamic weight allocation and selective scanning mechanisms, while the CA attention mechanism enhances the expressiveness of features and the effectiveness of the attention mechanism by embedding coordinate information. The combination of the two can complement each other and improve the overall feature extraction and target detection effects. The context-awareness provided by VSSBlock enables the CA attention mechanism to more accurately identify important feature areas when allocating attention, thereby improving the effectiveness of the attention mechanism. Therefore, the two modules work together to effectively improve the performance of the model.

[0071] Further configuration of this embodiment is as follows: Figure 1 and Figure 4 As shown in FIG. 1 , in order to reduce the amount of calculation of the model, the present invention designs a lightweight detection head and constructs a lightweight rotating frame detection head based on the lightweight detection head, as shown in FIG. Figure 4 As shown in (a).

[0072] Specifically, this embodiment improves the original YOLOv8 detection head, adopts group convolution instead of standard convolution in the classification head, and sets the number of groups to one eighth of the number of input channels, so that a lightweight detection head is obtained.

[0073] Furthermore, based on the lightweight detection head, an angle branch with an output channel number of 1 is added, and the output is mapped to the required angle range in combination with the sigmoid activation function. Finally, the output is composed of the regression information and category confidence, thereby constructing a lightweight rotation box detection head for rotating targets.

[0074] It should be noted that the first two convolutions are used to extract features, and the last two-dimensional convolution is used to change the number of channels. The difference is that the number of output channels is different. The angle branch only needs to predict one angle information, so the number of output channels is 1. The number of channels of the classification branch and the regression branch is 16. In addition, after obtaining the output, it is necessary to perform some operations on each branch to obtain the final result, such as the specific angle or the coordinates of the object. This operation is a conventional mathematical method, which is not specifically described in this embodiment.

[0075] Specifically, the schematic diagram of group convolution is shown in the attached Figure 4 As shown in (b). Figure 4 The standard convolution is introduced above (b). As can be seen from the figure, in the standard convolution, the entire input data will be convolved together. The input data is: H×W×c1, and the convolution kernel size is K×K, the channel is c1, and there are a total of c2 convolution kernels. Assume that the output resolution remains unchanged. Figure 4 The group convolution introduced in (b) is to divide the data into g groups and perform convolution separately (divided into two groups in the figure). This grouping is divided in depth, that is, every c1 / g channels are grouped into one group. Correspondingly, the depth of each group of convolution kernels also becomes one-g of the number of input channels, that is, c1 / g, and the number of convolution kernels in each group becomes one-g of the number of output channels c2, that is, c2 / g. Then, the convolution kernels of each group are convolved with the input data in their corresponding groups. After obtaining the output data, they are combined in a splicing manner. The final number of output channels is still c2. That is, after grouping, g identical convolution processes will be operated in parallel. The input data of each process is H×W×c1 / g, the convolution kernel size is K×K×c1 / g, there are a total of c2 / g, and the output data is H×W×c2 / g.

[0076] It is understandable that the computational complexity of the detection head in the YOLOv8 model accounts for more than 40% of the total computational complexity, so lightweighting the detection head can achieve good results. Currently, most improvements to lightweighting the detection head are to share the standard convolution of the classification head and regression head, and use lightweight convolution instead of standard convolution. In the YOLOv10 paper, the impact of classification errors and regression errors was analyzed, and it was found that classification errors have less impact on the performance of the model. Therefore, in this embodiment, by only lightweighting the classification head, the overhead of the detection head can be reduced with little loss in model performance.

[0077] This embodiment further provides the following comparison of the number of parameters and the amount of calculation between standard convolution and grouped convolution.

[0078] Table 1 Comparison between standard convolution and grouped convolution

[0079]

[0080] From the comparison in Table 1, we can see that compared with standard convolution, the amount of computation and parameters of group convolution is 1 / g of that of standard convolution.

[0081] In order to improve the detection effect of the lightweight model, the present invention adopts a new IOU calculation method MPDIoU. It is a new bounding box similarity metric based on the minimum point distance, which includes all relevant factors considered in the existing loss function, namely, overlapping or non-overlapping areas, center point distance, and width and height deviations, while simplifying the calculation process.

[0082] It can be understood that, in general, a unique rectangle can be defined using the coordinates of the upper left corner and the lower right corner. MPDIoU is designed based on this geometric property to minimize the distance between the upper left corner and the lower right corner of the predicted bounding box and the true bounding box. The algorithm of MPDIoU is as follows:

[0083] Table 2 MPDIoU algorithm

[0084]

[0085] It can be understood that MPDIoU simplifies the similarity comparison between two bounding boxes and can adapt to overlapping or non-overlapping bounding box regression. At the same time, using MPDIoU, faster convergence speed and more accurate regression results are achieved.

[0086] Step 4: Use the preprocessed dataset to train and verify the MCD-YOLO network model.

[0087] Specifically, in this embodiment, the present invention divides the data set into a training set and a validation set according to a ratio of 7:3, sets the training coefficient to 400 rounds, and inputs 8 pictures for each training. During the training process, wandb is used to observe the training process in real time. After the training is completed, the trained weights are saved.

[0088] Step 5: Use the remote sensing image to be detected as input to perform target detection through the pre-trained MCD-YOLO network model.

[0089] Specifically, in this embodiment, the output feature maps of the second C2f module, the first C2f_LVMB module, and the SPPF_CA module in the backbone network of the MCD-YOLO network model are used as inputs of the neck network, and the output of the neck network is used as inputs of the head network.

[0090] Finally, this embodiment illustrates the effect achieved by this invention with the help of the accompanying drawings and data. In order to further test the effect of the model, the benchmark model YOLOv8n-obb uses the obb detection head officially provided by ultralytics through ablation test. The experimental results are shown in Table 1, where CA represents the SPPF_CA module. The model with the best detection effect proposed by the present invention is YOLOv8-obb+C2f_LVMB+CA. Compared with the original model, AP 50 The value increased by 1.7 percentage points, AP 50-95 In order to further balance real-time performance and high precision, the present invention has made further improvements on the basis of this model and proposed the MCD-YOLO model, which reduces the amount of calculation by 27% and the number of parameters by 6% compared to YOLOv8-obb. 50 The value increased by 0.9 percentage points, AP 50-95 An increase of 0.8 percentage points.

[0091] Table 3. Comparison of ablation experiments

[0092]

[0093] In order to demonstrate the effect achieved by the invention, Figure 5-Figure 7 Explanation: Compare the detection image of the improved model with the original model. The left side is the detection image of the original model, and the right side is the detection image of the improved model. Figure 5 As shown in the figure, the original model identified the object at the edge as both an ordinary airplane and a helicopter, while the improved model correctly identified it as an ordinary airplane. Figure 6 As shown in Figure 2, when the characteristics of the football field are not obvious, the original model will miss the detection, while the improved model can accurately identify the football field; as shown in Figure 2 Figure 7 As shown in Figure 2, the original model mistakenly identified the small vehicle as a ship, while the improved model can correctly identify it. Figure 5 and Figure 7 ,It can be seen that the detection performance of the MCD-YOLO model is better than the original model under ,objects of different sizes and different backgrounds.

[0094] The embodiments of the present invention are described in detail above with reference to the accompanying drawings, but the present invention is not limited to the described embodiments. For those skilled in the art, various changes, modifications, substitutions and variations of these embodiments including components are made without departing from the principles and spirit of the present invention, and still fall within the scope of protection of the present invention.

Claims

1. A remote sensing image target detection method based on MCD-YOLO network, characterized in that: The steps include: Step 1: Obtain and preprocess the public remote sensing dataset; Step 2: Configure the training environment; Step 3, construct an MCD-YOLO network model, wherein the MCD-YOLO network model uses the YOLOv8 network as the basic network, replaces the C2f module in the backbone network of the YOLOv8 network with the C2f_LVMB module as the feature extraction module, and integrates the CA attention mechanism into the SPPF module in the backbone network; replaces the original detection head with a lightweight rotating box detection head in the head network of the YOLOv8 network; The original YOLOv8 network backbone network is provided with 4 C2f modules, and the latter two are replaced by C2f_LVMB modules; the C2f_LVMB module uses VSSBlock module to replace the Bottleneck module in the C2f module; The method for extracting features by the C2f_LVMB module is: First, use 1*1 convolution to change the number of channels i of the input feature map to the number of output channels m specified in the model configuration file. Then use the Split function to split the input into two branches with a channel number of 0.5*m. One branch extracts image features through n VSSBlock modules, and then connects it with the other branch through n residual connections to obtain a feature map with a channel number of (n+2)*0.5m. Finally, use 1*1 convolution to change the channel number to m to get the output. The method for constructing the lightweight rotating box detection head is as follows: group convolution is used to replace the standard convolution in the original detection head to obtain a lightweight detection head, and then an angle branch with an output channel number of 1 is added on the basis of the lightweight detection head, and the output is mapped to the required angle range in combination with the sigmoid activation function, and finally the output is composed of regression information and category confidence, thereby constructing a lightweight rotating box detection head; The MCD-YOLO network model uses MPDIoU as the loss function, and the algorithm is as follows: Define two arbitrary convex traits: Input image width and height: w, h; For A and B, (x1 A ,y1 A ), (x2 A ,y2 A ) represents the coordinates of the upper left corner and lower right corner of A; (x1 B ,y1 B ), (x2 B , y2 B ) represents the coordinates of the upper left corner and lower right corner of B; d1 2 =(x1 B -x1 A ) 2 +(y1 B -y1 A ) 2 ; <h2 style=";text-align:left;direction:ltr">d2<h2 style=";text-align:left;direction:ltr"> 2 <h2 style=";text-align:left;direction:ltr"> (x2)<h2 style=";text-align:left;direction:ltr"> B <h2 style=";text-align:left;direction:ltr"> -x2<h2 style=";text-align:left;direction:ltr"> A <h2 style=";text-align:left;direction:ltr"> )<h2 style=";text-align:left;direction:ltr"> 2 <h2 style=";text-align:left;direction:ltr"> +(y2<h2 style=";text-align:left;direction:ltr"> B <h2 style=";text-align:left;direction:ltr"> -y2<h2 style=";text-align:left;direction:ltr"> A <h2 style=";text-align:left;direction:ltr"> )<h2 style=";text-align:left;direction:ltr"> 2 <h2 style=";text-align:left;direction:ltr"> ; MPDIoU=(A∩B) / (AUB)-d1 2 / (w 2 +h 2 )-d2 2 / (w 2 +h 2 ); Step 4: Use the preprocessed data set to train and verify the MCD-YOLO network model; Step 5: Use the remote sensing image to be detected as input to perform target detection through the pre-trained MCD-YOLO network model.

2. The remote sensing image target detection method based on the MCD-YOLO network according to claim 1, characterized in that: The preprocessing method is to process the high-resolution image by segmentation, thereby dividing the image into blocks of fixed size.

3. The remote sensing image target detection method based on the MCD-YOLO network according to claim 2, characterized in that: The number of groups g of the grouped convolution is set to one eighth of the number of input channels.

4. The remote sensing image target detection method based on the MCD-YOLO network according to claim 3, characterized in that: The grouped convolution divides the input data with c1 channels into g groups for separate convolutions. The number of channels in each group is one-g of the input channels, i.e. c1 / g. The depth of each group of convolution kernels also becomes one-g of the input channels, i.e. c1 / g. The number of convolution kernels in each group becomes one-g of the output channels, c2. The convolution kernels of each group are then convolved with the input data in their corresponding groups. After obtaining the output data, they are combined in a splicing manner. The final number of output channels is still c2.

5. The remote sensing image target detection method based on the MCD-YOLO network according to claim 1, characterized in that: The training method of the MCD-YOLO network model is as follows: the data set is divided into a training set and a validation set according to a ratio of 7:3, the training coefficient is set to 400 rounds, 8 pictures are input for each training, the training process is observed in real time through wandb during the training process, and the trained weights are saved after the training is completed.

Images