Multi-scale feature cross distillation method for remote sensing image change detection

Abstract

The invention provides a remote sensing image change detection-oriented multi-scale feature cross distillation method, which comprises the following steps of: scaling an obtained original remote sensing image according to a preset scale factor to obtain M input images with different scales and corresponding classification labels, and forming an input image training set; training the detection model by using the input image training set; the detection model adopts a parallel multi-branch detection structure, and each detection branch correspondingly receives an input image for obtaining an anchor frame proposal feature map and a classification result; calculating distillation loss based on the anchor frame proposal feature map of each branch; calculating the detection loss based on the classification result of each branch, and constructing a loss function in combination with the distillation loss and the detection loss to obtain a trained detection model; and inputting a to-be-detected original remote sensing image into the trained detection model, and outputting a classification result. The method has strong adaptability to scale change in the image, and improves the performance of single-scale image reasoning while ensuring the reasoning speed.

Description

Technical Field

[0001] The present invention belongs to the technical field of remote sensing image detection, and relates to a multi-scale feature cross-distillation method for remote sensing image change detection. Background Art

[0002] With the rapid development of space technology, remote sensing image-based detection has been widely applied. However, there may be huge inter-class scale differences in the ground object information in remote sensing images. To address this problem, existing detection technologies mainly adopt two types of methods: network-level and data-level methods:

[0003] Network-level methods focus on constructing novel network architectures to extract multi-scale features that are robust to scale changes, mainly including feature pyramid architectures and multi-branch architectures; data-level methods are dedicated to designing data augmentation strategies independent of the network architecture. The most commonly used data-level method for scale changes is multi-scale training and testing, which enhances the input image by adjusting the size of the input image at different resolutions.

[0004] However, the existing network-level methods have an insufficient improvement effect on the scale difference problem of remote sensing images. Multi-scale training and testing can only improve the model performance within a specific scale range, and since data augmentation is required during testing, the inference speed is severely reduced.

[0005] Therefore, how to provide a multi-scale feature cross-distillation method for remote sensing image change detection that can improve the detection performance while ensuring the inference speed is an urgent problem to be solved by those skilled in the art. Summary of the Invention

[0006] In view of this, the present invention proposes a multi-scale feature cross-distillation method for remote sensing image change detection, which has a strong adaptability to scale changes in images, and while ensuring the inference speed, also improves the performance of single-scale image inference.

[0007] To achieve the above object, the present invention adopts the following technical solutions:

[0008] The present invention discloses a multi-scale feature cross-distillation method for remote sensing image change detection, including the following steps:

[0009] Model training step:

[0010] Obtain the original remote sensing image, including the original remote sensing image and the corresponding classification labels, and scale the original remote sensing image according to a preset scale factor to obtain M different-scale input images and the corresponding classification labels, where M>1, to form an input image training set; the M different-scale input images include one original remote sensing image and M-1 scaled images;

[0011] Train the detection model using the input image training set; the detection model adopts a parallel M-branch detection structure, and each detection branch correspondingly receives an input image and is used to obtain an anchor box proposal feature map and a classification result;

[0012] Calculate the distillation loss based on the anchor box proposal feature maps of each branch; calculate the detection loss based on the classification results of each branch, and construct a loss function by combining the distillation loss and the detection loss, which is used to update the parameters of the detection model during the training process, and finally obtain the trained detection model; only the original remote sensing image detection branch in the parallel M-branch detection structure is retained in the trained detection model;

[0013] Model detection steps:

[0014] Input the original remote sensing image to be detected into the trained detection model, and output the classification result.

[0015] Preferably, in the model training step:

[0016] The preset scale factor is expressed as The scale factor sizes include: 0 < s m < 1, that is, downsample the original image; and s m > 1, that is, upsample the image.

[0017] Preferably, the structures of the detection branches in the parallel M-branch detection structure are the same.

[0018] Preferably, the detection branch structure includes a feature extraction network, a region proposal network, and a prediction head, and each detection branch shares the same weight;

[0019] The feature extraction network is used to extract a feature map from the input image;

[0020] The region proposal network is used to generate a number of anchor boxes on the feature map to obtain the anchor box proposal feature map;

[0021] The prediction head is used to perform classification prediction on the anchor box proposal feature map based on a deep learning algorithm to obtain the classification result.

[0022] Preferably, the calculation steps of the distillation loss include:

[0023] Calculate the spatial similarity between the anchor box proposal feature map generated by the original remote sensing image detection branch and the anchor box proposal feature map generated by the scaled image detection branch to obtain the distillation loss.

[0024] Preferably, the calculation steps of the detection loss include:

[0025] Compare the classification results generated by the original remote sensing image detection branch with the classification results generated by the scaled image detection branch to obtain the detection loss.

[0026] Preferably, in the step of constructing the loss function: introduce a distillation weight, calculate the product of the distillation weight and the distillation loss, sum the product and the detection loss to obtain the loss function, and update the detection model parameters based on the loss function in combination with the gradient descent method.

[0027] The present invention also discloses a multi-scale feature cross-distillation system for remote sensing image change detection according to the multi-scale feature cross-distillation method for remote sensing image change detection, including:

[0028] An image scaling module for obtaining an original remote sensing image, including the original remote sensing image and corresponding classification labels, and scaling the original remote sensing image according to a preset scale factor to obtain M corresponding input images of different scales and corresponding classification labels, where M>1, constituting an input image training set; the M input images of different scales include one original remote sensing image and M-1 scaled images;

[0029] A model construction module for constructing a detection model, the detection model adopting a parallel M-branch detection structure, and each detection branch correspondingly receiving an input image and being used to obtain an anchor box proposal feature map and a classification result;

[0030] A model training module for training the detection model using the input image training set; calculating the distillation loss based on the anchor box proposal feature maps of each branch; calculating the detection loss based on the classification results of each branch, and constructing a loss function by combining the distillation loss and the detection loss for updating the detection model parameters during the training process, and finally obtaining the trained detection model; only the original remote sensing image detection branch in the parallel M-branch detection structure is retained in the trained detection model;

[0031] A model detection module for inputting the original remote sensing image to be detected into the trained detection model and outputting a classification result.

[0032] The present invention also discloses an electronic device, including a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that when the processor executes the program, the steps of the multi-scale feature cross-distillation method for remote sensing image change detection are implemented.

[0033] The present invention also discloses a computer-readable storage medium, on which a program is stored, and when the program is executed by a processor, the steps of the multi-scale feature cross-distillation method for remote sensing image change detection are implemented.

[0034] As can be seen from the above technical solutions, compared with the prior art, the beneficial effects of the present invention include:

[0035] The present invention utilizes a multi-scale multi-branch structure and an adaptive cross-scale knowledge distillation method, aiming to adaptively integrate multi-scale information to improve the performance of single-scale inference. The same network structure and shared weights in the multi-branch structure enable the model to have strong adaptability to scale changes in images after training, and can achieve accuracy equivalent to or higher than that of the multi-scale testing method while maintaining the single-scale inference speed. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are only the embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained according to the provided drawings;

[0037] Figure 1 is a flowchart of the multi-scale feature cross-distillation method for remote sensing image change detection provided by the embodiment of the present invention;

[0038] Figure 2 is a schematic diagram of the principle of the parallel M-branch detection structure provided by the embodiment of the present invention;

[0039] Figure 3 is a schematic diagram of the calculation principle of the adaptive cross-scale cross-distillation loss function provided by the embodiment of the present invention;

[0040] Figure 4 is an organizational block diagram of the multi-scale feature cross-distillation system for remote sensing image change detection provided by the embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0041] The following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present invention.

[0042] As Figure 1 shown, the first aspect of the embodiment of the present invention provides a multi-scale feature cross-distillation method for remote sensing image change detection, including the following steps:

[0043] Model training step:

[0044] Obtain the original remote sensing image, including the original remote sensing image and the corresponding classification labels, and scale the original remote sensing image according to a preset scale factor to obtain M different-scale input images and the corresponding classification labels, where M > 1, which constitute the input image training set, thereby changing the receptive field of the image by the same network structure; the M different-scale input images include one original remote sensing image and M - 1 scaled images;

[0045] Use the input image training set to train the detection model; the detection model adopts a parallel multi-branch detection structure, and the number of detection branches of the parallel multi-branch detection structure is the same as the number of input images. Each detection branch correspondingly receives one input image and is used to obtain the anchor box proposal feature map and the classification result;

[0046] Calculate the distillation loss based on the anchor box proposal feature maps of each branch; calculate the detection loss based on the classification results of each branch, and construct a loss function by combining the distillation loss and the detection loss to update the parameters of the detection model during the training process. Use the image training set to train the detection model so that the detection model achieves a good fitting effect on the training data, and finally obtain the trained detection model; only the original remote sensing image detection branch in the parallel M-branch detection structure is retained in the trained detection model;

[0047] Model detection steps:

[0048] Input the original remote sensing image to be detected into the trained detection model and output the classification result.

[0049] In one embodiment, the input image in the model training step can be expressed as where C, H, and W respectively represent the number of channels, width, and height of the input image. After scaling by the scale factor, I0 generates the above-mentioned different-scale input image pairs.

[0050] In this embodiment, the interesting ground objects in the input image are labeled, and the labeling information includes their positions and class information, which is used as the training set.

[0051] In one embodiment, before the input image pair is input into the parallel multi-branch structure in the model training step, it needs to be randomly flipped to increase the data diversity. Let the random flipping operation of the image be The result after random flipping can be expressed as

[0052] In one embodiment, in the model training step, the original image is scaled according to a preset scale factor to obtain input images of different scales for this image:

[0053] The preset scale factor is expressed as Where M represents the length of the scale factor, that is, each scale factor corresponds to a detection branch. The sizes of the scale factors include: 0 < s m < 1, which means downsampling the original image; and s m > 1, which means upsampling the image; the input image with a scale factor equal to 1 is the original remote sensing image, and the corresponding detection branch is the original remote sensing image detection branch.

[0054] In one embodiment, the structures of the detection branches in the parallel M-branch detection structure are all the same.

[0055] In one embodiment, the detection branch structure includes a feature extraction network, a region proposal network, and a prediction head, and each detection branch shares the same weights;

[0056] The feature extraction network is used to extract deep features from the input image to obtain a feature map;

[0057] The region proposal network is used to generate a number of anchor boxes on the feature map to obtain an anchor box proposal feature map;

[0058] The prediction head is used to perform classification prediction on the anchor box proposal feature map based on a deep learning algorithm to obtain a classification result.

[0059] In this embodiment, the feature extraction network is a feature pyramid structure composed of a deep neural network, which can fuse multi-scale features of the input image. This feature extraction network can be expressed as The features extracted from the input image by the feature extraction network can be expressed as

[0060] Those skilled in the art can understand that many feature extraction networks disclosed in the prior art can be used to form the feature extraction network, such as ResNET, VGGNet, etc. Combining these backbone networks into a feature pyramid structure can be used as the feature extraction network in this embodiment;

[0061] In this embodiment, the region proposal network will generate a large number of anchor boxes on the feature map, and determine positive region proposals and negative region proposals according to the intersection over union between the anchor boxes and the ground truth. This region proposal network can be expressed as All region proposal results can be expressed as

[0062] In this embodiment, the prediction head can adopt various common deep learning structures. For example, a prediction head composed of multiple fully connected layers is used to output prediction values of object category information and position information. The prediction head can be expressed as mapping the region proposal result to the original feature map of this branch, and finally inputting it into the prediction head. The prediction result can be expressed as

[0063] Those skilled in the art can understand that many prediction heads disclosed in the prior art can be used as the prediction head in the detection branch structure, and those skilled in the art can make a choice. The present invention does not specifically limit the type of the prediction head.

[0064] As Figure 2 shown, the specific execution process of this embodiment is as follows:

[0065] Input image I m For the corresponding branch, the number of m is determined by the number of scale factors. For the convenience of display in this embodiment, only one group of I is selected. m For the branch, all inputs other than the original image I0 to the corresponding branch are called other branches;

[0066] Each other detection branch is connected to the original remote sensing image detection branch through an adaptive cross-scale cross-distillation module, which can help the original branch learn the knowledge at the corresponding scale of other detection branches;

[0067] Before the image is input into the feature extraction network, random flipping processing needs to be performed to improve the diversity of data, which is expressed as For any branch, it is

[0068] The image features extracted from the preprocessed image data by the original branch feature extraction network can be expressed as F m , and the image features extracted from other branches can be expressed as F0;

[0069] The features extracted by the feature extraction network are input into the region proposal network. The region proposal network will generate a large number of anchor boxes on the feature map, and determine the positive region proposals and negative region proposals according to the intersection over union between the anchor boxes and the ground truth. This region proposal network can be expressed as The region proposal result can be expressed as

[0070] The feature extraction network will output positive proposal features. For the original branch, its output is For other branches, its output is where b i ∈[0, 1] 5 represents the parameters of the object rotation detection box after normalization;

[0071] The features output by the region proposal network need to be remapped back to the image features of the corresponding branch. For the original branch, the remapped result can be expressed as For other branches, the remapped result is

[0072] The prediction head can adopt the structure in various publicly available methods, and input the mapped result into the prediction head. The prediction result can be expressed as

[0073] Each branch will output the class information and location information of the detection result. The location information is marked in the form of a horizontal box. For any object in any image, it can be represented by y i to represent its class information and location information, where c i ∈{0,1} Nc , N c is the number of labeled classes in the dataset, and c i is the one-hot encoding for the labeled class. Then the output results of each branch are N p is the number of positive proposals output by the region detection network;

[0074] The positive value of the input object used in the training process of the detection model can be expressed as Based on the positive value and the predicted value, the loss during the training process can be calculated. The loss of the original branch is expressed as The loss of other branches is expressed as

[0075] In one embodiment, the calculation steps of the distillation loss include:

[0076] Calculate the spatial similarity between the anchor box proposal feature map generated by the original remote sensing image detection branch and the anchor box proposal feature map generated by the scaled image detection branch to obtain the distillation loss.

[0077] Specifically, during execution, the distillation loss is based on the positive and negative proposal features output by the original branch region proposal network, maps them to the other branches respectively, and compares them with the corresponding positive and negative proposal features of the other branches to obtain the loss value.

[0078] In this embodiment, the detection loss of the model can adopt the standard detection loss of a two-stage detector.

[0079] In one embodiment, the calculation steps of the detection loss include:

[0080] Compare the classification result generated by the original remote sensing image detection branch with the classification result generated by the scaled image detection branch to obtain the detection loss.

[0081] In one embodiment, in the step of constructing the loss function: to improve the cross-scale learning ability of the model and control the influence of the distillation loss on the model loss, a distillation weight is introduced to enable the model to adaptively adjust the influence of the distillation weight on the overall loss function. The product of the distillation weight and the distillation loss is calculated, and the sum of the product and the detection loss is used to obtain the loss function. The parameters of the detection model are updated based on the loss function in combination with the gradient descent method.

[0082] In this embodiment, at the initial stage of model training, the robustness of the detector is relatively low, and the distillation loss of the model is of relatively low quality. In the first few steps of training, the distillation weight will be set to 0, that is, the distillation loss in this case will be discarded when calculating the overall loss function.

[0083] As Figure 3 shown, the specific execution process of adaptive cross-scale cross-distillation is as follows:

[0084] Based on the parameters w i and α, the distillation loss is calculated. The distillation loss is:

[0085]

[0086] In the formula, N pos and N neg respectively represent the number of positive proposals and negative proposals output by the region proposal network, and M represents the total number of branches in the multi-branch structure;

[0087] For the original branch, is the feature of the positive proposal output by the region proposal network. After mapping this feature back to the feature map output by the feature extraction network through RoI align, the result is represented as

[0088] In order to enable the model to learn multi-scale knowledge under single-scale inference, it is necessary to map to the corresponding branches under different scale factors and then map it to the output of the feature extraction network of the corresponding branches;

[0089] Since the data preprocessing transformations between the detection branches are different, to ensure consistency, the output result of the original branch is subjected to the same transformation as that in the corresponding branch, that is, the transformation result is The transformation method is:

[0090]

[0091] After mapping to the feature map in the corresponding branch, the result is

[0092] The L2 loss is used to describe The spatial similarity between features, and the result is the root mean square error between the two;

[0093] To prevent the negative proposals from dominating the gradient during training, in this embodiment, an adaptive weight w i is used to enable the model to dynamically select valuable information for distillation during training, discard low-quality information, and the calculation method is as follows:

[0094]

[0095] In the formula, represents a positive value;

[0096] represents the standard detection loss between the true value and the prediction, which can be used to evaluate the quality of different RoI features. At this point, high-quality RoIs can be refined and low-quality features can be discarded.

[0097] The detection loss of the multi-branch detection structure can be expressed as:

[0098]

[0099] represents the detection loss of the standard two-stage detector;

[0100] Considering that at the beginning of model training, the detector is often not robust enough and the distillation loss is often of low quality, only the detection loss is considered, and the overall loss of the model is:

[0101]

[0102] A distillation weight α is introduced to adaptively change the influence of the distillation loss on the overall loss, so that the model can learn more high-quality information. At this time, the overall loss of the model is:

[0103]

[0104] In one embodiment, when training the multi-branch structure, the condition for training convergence is: when the number of training times reaches the specified number, the training ends; or, when the loss function of training drops to the specified threshold and remains lower than the threshold, the training ends.

[0105] In one embodiment, when the detection model is tested, only the original remote sensing image to be detected is input, and the classification result of the object to be detected in the image is obtained through the original remote sensing image detection branch. The classification result output by the detection model includes: the position information and category information of the object.

[0106] The second aspect of the embodiments of the present invention provides a multi-scale feature cross-distillation system for remote sensing image change detection according to the multi-scale feature cross-distillation method for remote sensing image change detection provided in the first aspect of the embodiments, including:

[0107] An image scaling module, configured to obtain an original remote sensing image, including the original remote sensing image and corresponding classification labels, and scale the original remote sensing image according to a preset scale factor to obtain M corresponding input images of different scales and corresponding classification labels, where M>1, forming an input image training set; the M input images of different scales include one original remote sensing image and M-1 scaled images;

[0108] A model construction module, configured to construct a detection model, and the detection model adopts a parallel M-branch detection structure, and each detection branch is correspondingly configured to receive an input image and obtain an anchor box proposal feature map and a classification result;

[0109] A model training module, configured to train the detection model by using the input image training set; calculate a distillation loss based on the anchor box proposal feature maps of each branch; calculate a detection loss based on the classification results of each branch, and construct a loss function by combining the distillation loss and the detection loss, for updating the parameters of the detection model during the training process, and finally obtaining a trained detection model; only the original remote sensing image detection branch in the parallel M-branch detection structure is retained in the trained detection model;

[0110] A model detection module, configured to input the original remote sensing image to be detected into the trained detection model and output a classification result.

[0111] The third aspect of the embodiments of the present invention provides an electronic device, including a memory, a processor, and a computer program stored on the memory and executable on the processor, wherein the processor implements the steps of the multi-scale feature cross-distillation method for remote sensing image change detection provided in the first aspect of the embodiments when executing the program.

[0112] The electronic device may be a server. The computer device includes a processor, a memory, a network interface, and a database connected through a system bus. Among them,

[0113] The processor of the electronic device is configured to provide computing and control capabilities.

[0114] The memory of the electronic device includes a non-volatile storage medium and an internal memory.

[0115] The non-volatile storage medium stores an operating system, a computer program, and a database.

[0116] The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium.

[0117] The database of the electronic device is used to store sample data.

[0118] The network interface of the electronic device is used to communicate with an external terminal through a network connection.

[0119] In the fourth aspect of the embodiments of the present invention, there is provided a computer-readable storage medium, on which a program is stored. When the program is executed by a processor, the steps of the multi-scale feature cross-distillation method for remote sensing image change detection provided in the first aspect of the embodiment are implemented.

[0120] Those skilled in the art can understand that all or part of the processes in the above-described embodiment methods can be completed by instructing relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above-mentioned various methods. Among them,

[0121] Any reference to a memory, storage, database, or other medium used in the various embodiments provided in the present application may include non-volatile and / or volatile memories. The non-volatile memory may include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory.

[0122] The volatile memory may include random access memory (RAM) or an external cache memory. By way of illustration and not limitation, RAM is available in many forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDR SDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and Rambus dynamic RAM (RDRAM), etc.

[0123] The above has introduced in detail the multi-scale feature cross-distillation method for remote sensing image change detection provided by the present invention. In this embodiment, specific examples are used to elaborate on the principle and implementation manner of the present invention. The description of the above embodiments is only used to help understand the method and its core idea of the present invention; at the same time, for those of ordinary skill in the art, according to the idea of the present invention, there will be changes in the specific implementation manner and application scope. In summary, the content of this specification should not be construed as a limitation to the present invention.

[0124] The foregoing description of the disclosed embodiments enables those skilled in the art to implement or use the present invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to these embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A multi-scale feature cross-distillation method for remote sensing image change detection, characterized in that It includes the following steps: Model training step: Obtain the original remote sensing image, including the original remote sensing image and the corresponding classification labels, and scale the original remote sensing image according to a preset scale factor to obtain M different-scale input images and the corresponding classification labels, where M>1, forming an input image training set; The M different-scale input images include one original remote sensing image and M-1 scaled images; Use the input image training set to train the detection model; the detection model adopts a parallel M-branch detection structure, and each detection branch correspondingly receives one input image and is used to obtain an anchor box proposal feature map and a classification result; Calculate the distillation loss based on the anchor box proposal feature maps of each branch; Calculate the detection loss based on the classification results of each branch, combine the distillation loss and the detection loss to construct a loss function, which is used to update the parameters of the detection model during training, and finally obtain the trained detection model; only the original remote sensing image detection branch in the parallel M-branch detection structure is retained in the trained detection model; Model detection step: Input the original remote sensing image to be detected into the trained detection model and output the classification result.

2. The multi-scale feature cross-distillation method for remote sensing image change detection according to claim 1, characterized in that In the model training step: The preset scale factor is expressed as The size of the scale factor includes: 0 < S m < 1, that is, downsampling the original image; and S m > 1, that is, upsampling the image.

3. A multi-scale feature cross-distillation method for remote sensing image change detection according to claim 1, characterized in that The structures of the detection branches in the parallel M-branch detection structure are the same.

4. A multi-scale feature cross-distillation method for remote sensing image change detection according to claim 1, characterized in that, The detection branch structure includes a feature extraction network, a region proposal network, and a prediction head, and each detection branch shares the same weight; The feature extraction network is used to extract a feature map from the input image; The region proposal network is used to generate a number of anchor boxes on the feature map to obtain the anchor box proposal feature map; The prediction head is used to perform classification prediction on the anchor box proposal feature map based on a deep learning algorithm to obtain the classification result.

5. A multi-scale feature cross-distillation method for remote sensing image change detection according to claim 1, characterized in that The calculation steps of the distillation loss include: Calculate the spatial similarity between the anchor box proposal feature map generated by the original remote sensing image detection branch and the anchor box proposal feature map generated by the scaled image detection branch to obtain the distillation loss.

6. A multi-scale feature cross-distillation method for remote sensing image change detection according to claim 1, characterized in that The calculation steps of the detection loss include: Compare the classification result generated by the original remote sensing image detection branch with the classification result generated by the scaled image detection branch to obtain the detection loss.

7. A multi-scale feature cross-distillation method for remote sensing image change detection according to claim 1, characterized in that In the step of constructing the loss function: introduce a distillation weight, calculate the product of the distillation weight and the distillation loss, sum the product and the detection loss to obtain the loss function, and update the parameters of the detection model based on the loss function combined with the gradient descent method.

8. A multi-scale feature cross-distillation system for remote sensing image change detection according to the multi-scale feature cross-distillation method for remote sensing image change detection described in any one of claims 1-7, characterized in that, It includes: An image scaling module, which is used to obtain the original remote sensing image, including the original remote sensing image and the corresponding classification labels, and scale the original remote sensing image according to a preset scale factor to obtain M different-scale input images and the corresponding classification labels, where M>1, forming an input image training set; The M different-scale input images include one original remote sensing image and M-1 scaled images; A model construction module, which is used to construct a detection model, and the detection model adopts a parallel M-branch detection structure, and each detection branch correspondingly receives one input image and is used to obtain an anchor box proposal feature map and a classification result; A model training module, which is used to train a detection model by using the input image training set; calculate a distillation loss based on the anchor box proposal feature maps of each branch; calculate a detection loss based on the classification results of each branch, construct a loss function by combining the distillation loss and the detection loss, and use it to update the parameters of the detection model during the training process, and finally obtain a trained detection model; only the original remote sensing image detection branch in the parallel M-branch detection structure is retained in the trained detection model; A model detection module, which is used to input the original remote sensing image to be detected into the trained detection model and output a classification result.

9. An electronic device, comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of a multi-scale feature cross-distillation method for remote sensing image change detection according to any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, A program is stored thereon, and when the program is executed by a processor, it implements the steps of the multi-scale feature cross-distillation method for remote sensing image change detection according to any one of claims 1-7.

Images