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

Abstract

The application provides a multiscale feature cross-distillation method for remote sensing image change detection, comprising: scaling the obtained original remote sensing image according to a preset scale factor to obtain M input images of different scales and corresponding classification labels, and constructing an input image training set; training a detection model by using the input image training set; the detection model adopts a parallel multi-branch detection structure, each detection branch corresponds to receive an input image for obtaining an anchor box proposal feature map and a classification result; calculating a distillation loss based on the anchor box proposal feature map of each branch; calculating a detection loss based on the classification result of each branch, combining the distillation loss and the detection loss to construct a loss function, and obtaining a trained detection model; inputting a to-be-detected original remote sensing image into the trained detection model to output a classification result. The application has strong adaptability to scale changes in the image, and improves the performance of single-scale image reasoning while ensuring the reasoning speed.

Description

Technical Field

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

[0002] With the rapid development of space technology, detection based on remote sensing images has been widely applied. However, ground feature information in remote sensing images may exhibit significant inter-class scale differences. To address this issue, existing detection technologies mainly employ two types of methods: network-level and data-level methods.

[0003] Network-level methods focus on building novel network structures to extract multi-scale features that are robust to scale changes, mainly including feature pyramid architecture and multi-branch architecture; data-level methods are dedicated to designing data augmentation strategies that are independent of 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, existing network-level methods have not significantly improved the problem of scale differences in remote sensing images. Multi-scale training and testing can only improve model performance within a specific scale range, and the need for data augmentation during testing severely reduces inference speed.

[0005] Therefore, how to provide a multi-scale feature cross-distillation method for remote sensing image change detection that can ensure inference speed while improving detection performance is a problem that urgently needs 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 improves the performance of single-scale image inference while ensuring inference speed.

[0007] To achieve the above objectives, the present invention adopts the following technical solution:

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

[0009] Model training steps:

[0010] Acquire the original remote sensing image, including the original remote sensing image and the corresponding classification label, and scale the original remote sensing image according to a preset scale factor to obtain M input images of different scales and corresponding classification labels, M>1, which constitute the input image training set; the M input images of different scales include one original remote sensing image and M-1 scaled images.

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

[0012] Distillation loss is calculated based on the anchor box proposal feature maps of each branch; detection loss is calculated based on the classification results of each branch; a loss function is constructed by combining the distillation loss and the detection loss, which is used to update the detection model parameters during training, and finally the trained detection model is obtained; the trained detection model retains only the original remote sensing image detection branch in the parallel M-branch detection structure.

[0013] Model detection steps:

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

[0015] Preferably, in the model training step:

[0016] The preset scaling factor is expressed as: Scale factor sizes include: 0 m <1, meaning the original image is downsampled; and s m >1 means upsampling the image.

[0017] Preferably, the detection branches in the parallel M-branch detection structure all have the same structure.

[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 weights;

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

[0020] The region proposal network is used to generate several 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 for the distillation loss include:

[0023] The spatial similarity between the anchor frame proposal feature map generated by the detection branch of the original remote sensing image and the anchor frame proposal feature map generated by the scaling image detection branch is calculated to obtain the distillation loss.

[0024] Preferably, the calculation step of the detection loss includes:

[0025] ​The classification result generated by the detection branch of the original remote sensing image is compared with the classification result generated by the detection branch of the scaled image to obtain the detection loss.

[0026] Preferably, in the step of constructing the loss function: distillation weights are introduced, the product of the distillation weights and the distillation loss is calculated, the product is summed with the detection loss to obtain the loss function, and the detection model parameters are updated based on the loss function and the gradient descent method.

[0027] This invention also discloses a multi-scale feature cross-distillation system for remote sensing image change detection based on the aforementioned multi-scale feature cross-distillation method for remote sensing image change detection, comprising:

[0028] The image scaling module is used to acquire the original remote sensing image, including the original remote sensing image and the corresponding classification label, and to scale the original remote sensing image according to a preset scale factor to obtain M input images of different scales and corresponding classification labels, where M>1, which constitute the input image training set; the M input images of different scales include one original remote sensing image and M-1 scaled images.

[0029] The model building module is used to build a detection model. The detection model adopts a parallel M-branch detection structure. Each detection branch receives an input image and is used to obtain the anchor box proposal feature map and classification result.

[0030] The model training module is used to train the detection model using the input image training set; calculate the distillation loss based on the anchor box proposal feature map 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 detection model parameters during the training process, and finally obtain the trained detection model; the trained detection model retains only the original remote sensing image detection branch in the parallel M-branch detection structure.

[0031] The model detection module is used to input the original remote sensing image to be detected into the trained detection model and output the classification result.

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

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

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

[0035] This invention utilizes a multi-scale, multi-branch structure and an adaptive cross-scale knowledge distillation method to adaptively integrate multi-scale information, thereby improving the performance of single-scale inference. The identical network structure and shared weights in the multi-branch structure enable the model to possess strong adaptability to scale changes in images after training, achieving accuracy comparable to or higher than multi-scale testing methods while maintaining the speed of single-scale inference. Attached Figure Description

[0036] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

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

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

[0039] Figure 3 A schematic diagram illustrating the calculation principle of the adaptive cross-scale cross-distillation loss function provided in this embodiment of the invention;

[0040] Figure 4 This is a block diagram of a multi-scale feature cross-distillation system for remote sensing image change detection provided in an embodiment of the present invention. Detailed Implementation

[0041] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

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

[0043] Model training steps:

[0044] The original remote sensing image is acquired, including the original remote sensing image and the corresponding classification label. The original remote sensing image is scaled according to a preset scale factor to obtain M input images of different scales and corresponding classification labels, M>1, which constitute the input image training set. Then, the receptive field of the same network structure to the image is changed. The M input images of different scales include one original remote sensing image and M-1 scaled images.

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

[0046] Distillation loss is calculated based on the anchor box proposal feature maps of each branch; detection loss is calculated based on the classification results of each branch; a loss function is constructed by combining the distillation loss and the detection loss, which is used to update the detection model parameters during training; the detection model is trained using the image training set, so that the detection model achieves a good fit to the training data, and finally the trained detection model is obtained; the trained detection model retains only the original remote sensing image detection branch in the parallel M-branch detection structure.

[0047] Model detection steps:

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

[0049] In one embodiment, the input image in the model training step can be represented as: Where C, H, and W represent the number of channels, width, and height of the input image, respectively. I0 generates the input image pairs at different scales after being scaled by a scale factor.

[0050] In this embodiment, the objects of interest in the input image are labeled, and the labeling information includes their location and class information, which is used as a training set.

[0051] In one embodiment, during the model training step, the input image needs to be randomly flipped before being applied to the parallel multi-branch structure to increase data diversity. Let the random flipping operation on the image be... The result after random flipping can be represented as

[0052] In one embodiment, during the model training step, the original image is scaled according to a pre-defined scale factor to obtain input images at different scales for that image:

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

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

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

[0056] Feature extraction networks are used to extract depth features from input images to obtain feature maps;

[0057] The region proposal network is used to generate several anchor boxes on the feature map to obtain the 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 the 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 represented as follows: The features extracted from the input image by the feature extraction network can be represented as:

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

[0061] In this embodiment, the region proposal network generates a large number of anchor boxes on the feature map. Positive and negative region proposals are determined based on the intersection-union ratio (IUU) between the anchor boxes and ground truth values. This region proposal network can be represented as follows: The results of all regional proposals can be represented as follows:

[0062] In this embodiment, the prediction head can employ various commonly used deep learning structures, such as a prediction head composed of multiple fully connected layers, used to output predicted values ​​for object category and location information. The prediction head can be used... This means that the region proposal result is mapped onto the original feature map of this branch, and finally input into the prediction head. The prediction result can be represented as follows: ​

[0063] Those skilled in the art will understand that many prediction heads disclosed in the prior art can be used as prediction heads in the detection branch structure, and those skilled in the art can make selections. This invention does not impose specific restrictions on the types of prediction heads.

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

[0065] Input image I m For the corresponding branches, the number of m is determined by the number of scale factors. In this embodiment, for ease of demonstration, only one set of I is selected. m Branches: All branches corresponding to non-original image I0 inputs are called other branches;

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

[0067] Before the image is input into the feature extraction network, it needs to be randomly flipped to improve data diversity, as shown below. For any branch, it is...

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

[0069] The features extracted by the feature extraction network are input into the region proposal network. The region proposal network generates a large number of anchor boxes on the feature map. Positive and negative region proposals are determined based on the intersection-union ratio (IoU) between the anchor boxes and ground truth values. This region proposal network can be represented as... The regional proposal results can be represented as

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

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

[0072] The prediction head can adopt the structure of various publicly available methods. The mapped result is input into the prediction head, and the prediction result can be represented as follows:

[0073] Each branch will output the category and location information of the detection results. The location information is marked with a horizontal bounding box. For any feature in any image, it can be represented by y. i This indicates its category information and location information. Where c i ∈{0,1} Nc N c c is the number of labeled categories in the dataset. i If the one-hot encoding is used for the labeled categories, then the output results of each branch are: N p The number of positive proposals output by the regional detection network;

[0074] The positive values ​​of the input objects used in the training of the detection model can be represented as: The loss during training can be calculated based on the positive value and the predicted value. The loss of the original branch is expressed as follows: The loss of other branches is expressed as

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

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

[0077] In practice, the distillation loss is based on the positive and negative proposal features output by the original branch region proposal network. These features are mapped to the remaining branches, and the loss value is obtained by comparing them with the positive and negative proposal features corresponding to the remaining branches.

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

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

[0080] The classification results generated by the detection branch of the original remote sensing image are compared with the classification results generated by the detection branch of the scaled image to obtain the detection loss.

[0081] In one embodiment, the step of constructing the loss function is as follows: to improve the model's cross-scale learning ability and control the impact of distillation loss on the model loss, distillation weights are introduced so that the model can adaptively adjust the impact of distillation weights on the overall loss function. The product of distillation weights and distillation loss is calculated, and the product is summed with the detection loss to obtain the loss function. The detection model parameters are updated based on the loss function combined with the gradient descent method.

[0082] In this embodiment, during the initial stage of model training, the detector robustness is low and the model's distillation loss is of low quality. During the initial step of training, the distillation weight will be 0, meaning that the distillation loss in this case will be discarded when calculating the overall loss function.

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

[0084] Based on parameters w i α is used to calculate distillation loss. for:

[0085]

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

[0087] For the original branch, The positive proposal features output by the Region Proposal Network (RoI) are mapped back to the feature map output by the Feature Extraction Network using RoI align. The result is represented as follows:

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

[0089] Because the preprocessing transformations of the data differ between the various detection branches, to ensure consistency, the output results of the original branches are... Perform the same transformation as in the corresponding branch, i.e., the transformation result is... The transformation method is as follows:

[0090]

[0091] Will After mapping to the feature maps in the corresponding branches, the result is:

[0092] L2 loss is used to describe and The spatial similarity between features is expressed as the root mean square error between them.

[0093] To prevent negative proposals from dominating the gradient during training, this embodiment uses adaptive weights w. i The model can dynamically select valuable information for distillation and discard low-quality information during training. The calculation method is as follows:

[0094]

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

[0096] This represents the standard detection loss between the true value and the prediction, which can be used to evaluate the quality of different RoI features. This allows us to extract high-quality RoIs and discard low-quality features.

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

[0098]

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

[0100] Considering that detectors are often not robust enough at the beginning of model training, and distillation loss is often of low quality, we only consider detection loss. The overall model loss is:

[0101]

[0102] By introducing a distillation weight α to adaptively change the impact of distillation loss on the overall loss, the model can learn more high-quality information. The overall model loss is then:

[0103]

[0104] In one embodiment, when training the multi-branch structure, the training convergence condition is: training ends when the number of training iterations reaches a specified number; or training ends when the training loss function decreases to a specified threshold and remains below the threshold.

[0105] In one embodiment, during the testing of the detection model, only the original remote sensing image to be detected is input. The detection branch of the original remote sensing image obtains the classification result of the object to be detected in the image. The classification result output by the detection model includes: the location information and category information of the object.

[0106] A second aspect of this invention provides a multi-scale feature cross-distillation system for remote sensing image change detection, based on a first aspect of this invention, comprising:

[0107] The image scaling module is used to acquire the original remote sensing image, including the original remote sensing image and the corresponding classification label, and to scale the original remote sensing image according to a preset scale factor to obtain M input images of different scales and corresponding classification labels, where M>1, which constitute the input image training set; the M input images of different scales include one original remote sensing image and M-1 scaled images.

[0108] The model building module is used to build the detection model. The detection model adopts a parallel M-branch detection structure. Each detection branch receives an input image and is used to obtain the anchor box proposal feature map and classification result.

[0109] The model training module is used to train the detection model using the input image training set; it calculates the distillation loss based on the anchor box proposal feature maps of each branch; it calculates the detection loss based on the classification results of each branch, and constructs a loss function by combining the distillation loss and the detection loss, which is used to update the detection model parameters during the training process, and finally obtains the trained detection model; the trained detection model retains only the original remote sensing image detection branch in the parallel M-branch detection structure.

[0110] The model detection module is used to input the original remote sensing image to be detected into the trained detection model and output the classification result.

[0111] A third aspect of the present invention provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that the processor executes the program to implement the steps of a multi-scale feature cross-distillation method for remote sensing image change detection provided in the first aspect of the present invention.

[0112] The electronic device can be a server. This computer device includes a processor, memory, network interface, and database connected via a system bus.

[0113] The processor in an electronic device is used to provide computing and control capabilities.

[0114] The memory of electronic devices includes non-volatile storage media and internal memory.

[0115] Non-volatile storage media store operating systems, computer programs, and databases.

[0116] Internal memory provides the environment for the operation of operating systems and computer programs in non-volatile storage media.

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

[0118] The network interface of an electronic device is used to communicate with external terminals via a network connection.

[0119] A fourth aspect of the present invention provides a computer-readable storage medium having a program stored thereon, which, when executed by a 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 present invention.

[0120] Those skilled in the art will understand that implementing all or part of the processes in the above embodiments can be accomplished by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods.

[0121] Any references to memory, storage, database, or other media used in the embodiments provided in this application may include non-volatile and / or volatile memory. 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] Volatile memory may include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and RAMbus dynamic RAM (RDRAM), etc.

[0123] The multi-scale feature cross-distillation method for remote sensing image change detection provided by the present invention has been described in detail above. Specific examples have been used in this embodiment to illustrate the principle and implementation of the present invention. The description of the above embodiments is only for the purpose of helping to understand the method and core idea of ​​the present invention. At the same time, for those skilled in the art, there will be changes in the specific implementation and application scope based on the idea of ​​the present invention. Therefore, the content of this specification should not be construed as a limitation of the present invention.

[0124] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined in these embodiments may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the 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, The method comprises the following steps: The model training step: Obtain an original remote sensing image, including an original remote sensing image and a corresponding classification label, and scale the original remote sensing image according to a preset scale factor to obtain M input images of different scales and corresponding classification labels, M > 1, to form an input image training set; The M input images of different scales include one original remote sensing image and M-1 scaled images; The input image training set is used to train the detection model; the detection model adopts a parallel M-branch detection structure, each detection branch receives one input image and is used to obtain anchor box proposal feature maps and classification results; 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 feature maps from the input image; The region proposal network is used to generate a plurality of anchor boxes on the feature maps to obtain the anchor box proposal feature maps; The prediction head is used to classify and predict the anchor box proposal feature maps based on a deep learning algorithm to obtain classification results; The anchor box proposal feature maps of each branch are used to calculate a distillation loss; A detection loss is calculated based on the classification results of each branch, and a loss function is constructed by combining the distillation loss and the detection loss, which is used to update the detection model parameters in the training process to finally obtain a trained detection model; the trained detection model only retains the original remote sensing image detection branch in the parallel M-branch detection structure; The calculation steps of the distillation loss include: calculating the spatial similarity between the anchor box proposal feature maps generated by the original remote sensing image detection branch and the anchor box proposal feature maps generated by the scaled image detection branch to obtain the distillation loss; The calculation steps of the detection loss include: comparing 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; In the step of constructing the loss function: a distillation weight is introduced, the product of the distillation weight and the distillation loss is calculated, the sum of the product and the detection loss is the loss function, and the detection model parameters are updated based on the loss function and the gradient descent method; The model detection step: The original remote sensing image to be detected is input into the trained detection model, and the classification result is output.

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 represented as The scale factor size includes: 0 <1, i.e. down-sampling the original image; and >1, i.e. up-sampling the image.

3. The 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 system for remote sensing image change detection according to any one of claims 1-3, characterized in that, It comprises: An image scaling module is used to obtain an original remote sensing image, including an original remote sensing image and a corresponding classification label, and scale the original remote sensing image according to a preset scale factor to obtain M input images of different scales and corresponding classification labels, M > 1, to form an input image training set; The M input images of different scales include one original remote sensing image and M-1 scaled images; A model construction module is used to construct a detection model, and the detection model adopts a parallel M-branch detection structure, each detection branch receives one input image and is used to obtain anchor box proposal feature maps and classification results; A model training module is configured to train a detection model by using the input image training set; and calculate a distillation loss based on the anchor box proposal feature maps of each branch; A detection loss is calculated based on the classification results of each branch, and a loss function is constructed by combining the distillation loss and the detection loss, which is used to update the parameters of the detection model in the training process, and finally obtain a trained detection model; the trained detection model only retains the original remote sensing image detection branch in the parallel M-branch detection structure; A model detection module is configured to input a to-be-detected original remote sensing image into the trained detection model, and output a classification result.

5. An electronic device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, The processor executes the program to implement the steps of the multi-scale feature cross-distillation method for remote sensing image change detection according to any one of claims 1-3.

6. A computer-readable storage medium, characterized in that, A program is stored thereon, 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 according to any one of claims 1-3 are implemented.

Images