Environmental change detection system and method based on prior prompt dynamic adjustment network

Abstract

The invention provides an environment change detection system and method based on a prior prompt dynamic adjustment network, and the system is characterized in that an encoder module is configured to carry out the feature extraction and encoding of an input image, and obtains a fine-grained feature; taking the fine-grained features as embedding vectors of K and V; the fine-grained features are sent to a feature mining module, the feature mining module is configured to perform mining operation on the fine-grained features, obtain position information and content information and send the position information and the content information as prior prompts to a decoder module, and the decoder module is configured to give the prior prompts to a query to obtain a new query; based on the weight ratio of one-to-many label matching and one-to-one label matching, the new query and the K and V embedded vectors are subjected to first operation, an output vector is obtained, the new query is updated, and the change detection module is configured to perform second operation on the final output vector and mask information obtained based on fine-grained features, and detecting the change of the input image based on the result of the second operation.

Description

Technical Field

[0001] The present invention relates to the field of computer technology, and in particular to an environment change detection system and method for dynamically adjusting a network based on priori prompts. Background Art

[0002] With the rapid development of society and the acceleration of urbanization, the application of advanced remote sensing technology in monitoring urban changes has become a necessity. Remote sensing technology can quickly capture the physical state changes of buildings and infrastructure through comparative analysis of images of the same area at different times, provide key decision-making support for rescue teams, promote the precise allocation of rescue resources, and provide important planning basis for post-disaster reconstruction. Satellite remote sensing plays an important role in disaster monitoring with its advantages of regular coverage, wide-area monitoring and high-resolution imaging. Satellites such as the LandSat series, Ikonos, QuickBird, WorldView, and China's independently developed Gaofen-1 and Gaofen-2 satellites all provide solid data support for change detection. One of the main applications of remote sensing technology in disaster detection is change detection. By comparing and analyzing satellite images from different periods, the changes in buildings can be quickly determined and the damage caused to buildings by disasters can be captured. This technology not only helps to timely assess the impact range of disasters, but also provides an important basis for subsequent rescue operations. The change detection results provide key data support for rescue decisions, enable rescue resources to be accurately allocated to where they are most needed, and provide an important planning basis for post-disaster reconstruction. As a type of change detection, building change detection is important because it can quickly capture the damage of buildings, provide key decision-making support for rescue teams, and assist in post-disaster reconstruction. By identifying which buildings can be repaired and which need to be demolished and rebuilt, reconstruction efficiency can be improved. By using deep learning technology to analyze remote sensing images, the accuracy and efficiency of change detection can be improved through automatic learning and extraction of complex features, thus showing significant advantages in post-disaster building damage assessment.

[0003] In the field of building change detection and damage assessment, although traditional methods such as image difference, image ratio, image regression, and change vector analysis (CVA) have achieved certain results in certain application scenarios, they usually ignore the local information of the image, resulting in high sensitivity to noise and registration errors. In addition, these methods rely on manually designed fixed features and may not fully capture the complexity and diversity of changes. In contrast, deep learning methods represented by convolutional neural networks and transformers can automatically learn richer and more robust image representations through their powerful feature extraction capabilities, thus showing significant advantages in change detection or damage assessment tasks. Deep learning models can recognize and adapt to complex change patterns in images, including local details and global context, while being more robust to noise and registration errors. In addition, the end-to-end learning framework of deep learning models allows learning change features directly from raw pixel data without complex feature engineering, which further simplifies the process and improves performance. However, current deep learning methods do not work well when further requirements are met. For example, in the process of building change detection, current network models cannot simultaneously enhance the model representation capabilities and accelerate convergence. Summary of the invention

[0004] In order to solve the above problems, the present invention provides an environment change detection system and method for dynamically adjusting a network based on a priori prompts.

[0005] According to one aspect of the present invention, an environment change detection system based on a priori prompts and dynamically adjusting the network is provided, comprising an encoder module, a feature mining module, a decoder module and a change detection module, wherein the encoder module is configured to: extract and encode features of an input image to obtain fine-grained features; use the fine-grained features as K, V The embedding vector is sent to the decoder module; the fine-grained features are sent to the feature mining module; the feature mining module is configured to: perform mining operations on the fine-grained features to obtain location information and content information; the location information and content information are sent to the decoder module as prior hints; the decoder module is configured to: assign the prior hint to the query to obtain a new query; in each iteration of training, based on the weight ratio of one-to-many label matching and one-to-one label matching, the new query and the K, V embedding vector are subjected to a first operation to obtain an output vector and update the new query; when a predetermined condition is met, the iteration is stopped and the final output vector is obtained; the change detection module is configured to: perform a second operation on the final output vector and the mask information obtained based on the fine-grained features, and detect changes in the input image based on the result of the second operation, wherein the decoder module further includes a dynamic label assignment module, and the dynamic label assignment module is configured to: change the weight ratio of one-to-many label matching and one-to-one label matching in each iteration of training.

[0006] Furthermore, as the number of iterations increases, the weight ratio of one-to-many label matching and one-to-one label matching continues to decrease.

[0007] Furthermore, the predetermined condition is reaching a set maximum number of iterations.

[0008] Furthermore, extracting and encoding features of the input image to obtain fine-grained features includes: performing a backbone operation on the input image to obtain coarse-grained features; and encoding the coarse-grained features to obtain fine-grained features.

[0009] Furthermore, the mask information is obtained by upsampling the fine-grained features.

[0010] Further, the first operation is a cross-attention operation.

[0011] Further, the second operation is a dot product operation.

[0012] According to one aspect of the present invention, a detection method for an environment change detection system based on a priori prompt dynamic adjustment network is provided, comprising: extracting and encoding features of an input image through an encoder module to obtain fine-grained features; sending the fine-grained features as K, V embedding vectors to a decoder module; sending the fine-grained features to a feature mining module; mining the fine-grained features through the feature mining module to obtain location information and content information; sending the location information and content information as priori prompts to a decoder module; assigning the priori prompts to a query through the decoder module to obtain a new query; in each iteration of training, based on the weight ratio of one-to-many label matching and one-to-one label matching, performing a first operation on the new query and the K, V embedding vector to obtain an output vector and update the new query; when a predetermined condition is met, stopping the iteration and obtaining the final output vector; performing a second operation on the final output vector and the mask information obtained based on the fine-grained features through a change detection module, and detecting changes in the input image based on the result of the second operation, wherein the detection method further comprises: changing the weight ratio of one-to-many label matching and one-to-one label matching through a dynamic label assignment module included in the decoder module during each iteration of training.

[0013] According to one aspect of the present invention, an electronic device is provided, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, the steps of the detection method of the environment change detection system for dynamically adjusting the network based on prior prompts are implemented.

[0014] According to one aspect of the present invention, a non-transitory computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a processor, the steps of the detection method of the environment change detection system for dynamically adjusting the network based on prior prompts are implemented as described above.

[0015] According to the present invention, a system and method for detecting environmental changes based on dynamic adjustment of networks based on prior prompts are provided, wherein, compared with random initialization, a method for reinitializing queries is proposed, prompts are added to queries, prior knowledge is given to the queries, and the queries contain content and location information. In order to better serve query learning and make the location and content as accurate as possible, a classifier is added after the encoder to obtain a more fine-grained location and content, and a feature mining module is added before the classic classifier to better improve the accuracy of the classifier. This effectively solves the problem of false detection that is prone to occur in building change detection under remote sensing images; in addition, a dynamic candidate box quantity adjustment strategy is proposed, a function with epoch as an independent variable is constructed, and a one-to-many label allocation strategy and a one-to-one label allocation strategy are controlled by the number of training rounds to realize one-to-many label matching supervision and one-to-one label matching. One-to-many label matching can better enhance the learning representation ability of the model, and one-to-one label matching can accelerate model convergence, realize end-to-end detection, and simplify training. Through dynamic function adjustment, the difference in the number of buildings in different images is avoided, and the problem of low efficiency of feature learning of only one-to-one supervision mechanism is solved. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 It is a schematic diagram showing the overall network structure of the present invention.

[0017] Figure 2 is a schematic diagram of an environment change detection system for dynamically adjusting a network based on priori prompts according to an embodiment of the present invention.

[0018] Figure 3 is a schematic diagram of a dynamic label allocation module according to an embodiment of the present invention.

[0019] Figure 4 is a diagram illustrating the ablation experiment results of the feature mining module according to an embodiment of the present invention.

[0020] Figure 5 is a diagram of an electronic device according to an embodiment of the present invention. DETAILED DESCRIPTION

[0021] Embodiments according to the present invention will be described below with reference to the accompanying drawings.

[0022] The present invention is mainly to solve the problem of low performance in building change detection in the current remote sensing field. In the satellite images taken, buildings often show discrete distribution characteristics, and the complex background and foreground blending are relatively complex, and there are significant omissions and false calls for building foreground targets. To address this problem, a neural network model based on dynamic adjustment of feature deep mining is proposed, and the performance of the neural network is improved through dynamic adjustment factors and deep mining modules. The present invention aims to enhance the model representation capability, propose a deep feature mining module and a dynamic label allocation module, further solve the problem of low accuracy and recall rate of previous models for building change detection in remote sensing data sets, and improve the performance of target detection.

[0023] Figure 1 It is a schematic diagram showing the overall network structure of the present invention.

[0024] Reference Figure 1 , Figure 1 The upper left corner is the scene before the disaster (model input). Figure 1 The lower right corner is the scene diagram after the disaster (model output). The input image is processed (for example, the backbone network) to obtain coarse-grained features (coarse-grained features). Fine-grained features are obtained by processing the coarse-grained features. There are three types of operations on fine-grained features: the first is to use the fine-grained features for feature mining in the subsequent feature miner to obtain content information (content) and location information (location) for subsequent operations in the decoder (as prior information given to the query in the decoder to approximate the objective distribution of the real world truth); the second is to use the fine-grained features as K, V embedding vectors for subsequent operations in the decoder; the third is to operate the fine-grained features to obtain mask information (mask) for subsequent mask calculations. In this model, a dynamic adjustment factor μ is introduced to control the weight ratio of one-to-many label matching and one-to-one label matching, ensuring the learning ability of the model while accelerating the convergence of the model.

[0025] Figure 2 is a schematic diagram of an environment change detection system 100 for dynamically adjusting a network based on a priori prompts according to an embodiment of the present invention.

[0026] Reference Figure 2 The change detection system 100 includes an encoder module 110 , a feature mining module 120 , a decoder module 130 and a change detection module 140 .

[0027] In an embodiment, the encoder module 110 can be configured to extract and encode features of the input image to obtain fine-grained features; send the fine-grained features as K, V embedding vectors to the decoder module 130; and send the fine-grained features to the feature mining module 120.

[0028] In an embodiment, the feature mining module 120 may be configured to perform mining operations on fine-grained features to obtain location information and content information; and send the location information and content information to the decoder module 130 as a priori prompts.

[0029] According to the present invention, through the operation of the feature mining module 120, the fine-grained features are mined to obtain the position information and content information; the position information and content information are sent to the decoder module as a priori prompts. By adding the position information and content information to the query as a priori knowledge, it is different from the most traditional random initialization, which helps to further match the true value of the data set label and accelerate the model training and convergence.

[0030] In an embodiment, the decoder module 130 can be configured to assign a priori hints to the query to obtain a new query; during each iteration of training, based on the weight ratio of one-to-many label matching and one-to-one label matching (which can be regarded as a dynamic adjustment factor, which will be described in detail below), the new query and the K, V embedding vectors are subjected to a first operation to obtain an output vector and update the new query; when a predetermined condition is met, the iteration is stopped and the final output vector is obtained.

[0031] In an embodiment, the change detection module 140 may be configured to perform a second operation on the final output vector of the decoder module 130 and the mask information obtained based on the fine-grained features of the encoder module 110, and detect changes in the input image based on the result of the second operation.

[0032] According to the present invention, the predetermined condition may be that the iterative training reaches a set maximum number of iterations, but the present invention is not limited thereto. For example, the predetermined condition may be set to that the detection change performance (eg, F1 mentioned below) reaches a predetermined threshold.

[0033] According to the present invention, extracting and encoding features of an input image to obtain fine-grained features may include: performing a backbone operation on the input image to obtain coarse-grained features; encoding the coarse-grained features (for example, through an L-layer encoder) to obtain fine-grained features.

[0034] According to the present invention, the mask information may be obtained by upsampling and restoring the fine-grained features.

[0035] According to the present invention, the first operation may be a cross-attention operation, but the present invention is not limited thereto.

[0036] According to the present invention, the second operation is a dot product operation, but the present invention is not limited thereto.

[0037] According to the present invention, the mining operation may be an MLP-based operation, but the present invention is not limited thereto.

[0038] In an embodiment, the decoder module 130 may further include a dynamic label assignment module 130 - 1 , which may be configured to change the weight ratio of one-to-many label matching and one-to-one label matching during each iteration of training.

[0039] The following reference Figure 3 The operation of the dynamic label allocation module 130 - 1 is described in detail.

[0040] In the present invention, one-to-many label matching and one-to-one label matching are introduced, and the one-to-many paradigm supervises the one-to-one paradigm, so that the model has both feature expression ability and model fast convergence ability. At this time, the input to the decoder is the query that obtains prior knowledge, which can be called a new query. The new query and the fine-grained features (such as the K, V embedding vectors above) are operated with attention.

[0041] Among them, one-to-many label matching means that there are multiple predicted boxes in the neural network that match one real box, and finally the most suitable candidate box is selected through the non-maximum suppression method. One-to-one label matching means that there is only one predicted box in the neural network that matches one real box.

[0042] The advantage of the one-to-many label matching paradigm is that the model has strong learning representation capabilities, but the disadvantage is that a large number of anchor frames need to be manually designed in the early stage of the model, the parameters are redundant, and it is affected by manual design; the advantage of the one-to-one label matching paradigm is that the model realizes end-to-end detection, the training process is simpler, and the model is easy to converge. The disadvantage is that only one candidate box is used to make judgments, the model's expression ability is reduced, and confusion is prone to occur in the reasoning process.

[0043] The present invention adopts a method combining one-to-many label matching and one-to-one label matching, i.e. one-to-many supervised one-to-one label matching. That is, in the early stage of model training, in order to maintain the model's feature expression learning ability, one-to-many supervision is adopted. As the number of training rounds increases, the weight ratio of one-to-many label matching and one-to-one label matching is changed through a dynamic adjustment factor, so as to ensure the learning ability of the model while taking into account the acceleration of model convergence; in the later stage of model training, the learning ability of the neural network model is relatively saturated, and the dynamic adjustment factor is changed, so that the model tends to the traditional one-to-one label matching, which can accelerate the convergence of the neural network and make the training process more concise.

[0044] The calculation of the dynamic adjustment factor is as follows: (1) In the above formula (1), T is the total number of training rounds of the model, t is the current number of training rounds, and n is the initial weight ratio of one-to-many supervision and one-to-one supervision. The balanced label allocation strategy is implemented. That is, from the above equation, we can see that the dynamic adjustment factor The initial value of is n. As the iteration proceeds, The value keeps decreasing.

[0045] Thus, the dynamic label assignment module 130 - 1 of the decoder module 130 may be configured to change the weight ratio of one-to-many label matching and one-to-one label matching during each iteration of training.

[0046] Based on the above concept of dynamic balanced label allocation, the algorithm of dynamic balanced label allocation is shown below:

[0047] In order to verify the effect achieved by the technical solution of the present invention, the present invention uses the XBD data set for testing.

[0048] The XBD dataset is an innovative large-scale satellite image dataset designed to assess building damage after natural disasters. The dataset contains 22,068 1024x1024 pixel resolution satellite images from 19 different natural disaster events, covering a geographic area of ​​more than 45,000 square kilometers. The image data is collected from the WorldView-3 satellite with a spatial resolution of 0.3 meters per pixel. The XBD dataset introduces a four-level damage annotation scale, from no damage (0) to complete destruction (4), providing a standardized assessment tool for building damage assessment.

[0049] This test was completed using 2*4090 24G and 1 A100 40G. To facilitate testing, the dataset was converted to COCO for better multi-model detection. The model is based on the PyTorch framework and trained in the cloud. During the training process, the Stochastic Gradient Descent (SGD) optimizer was used, with an initial learning rate of 1e-2. As the training learning rate decayed to 2e-3, the total number of training rounds in the training process was 500 (epochs).

[0050] This test uses F1 as the evaluation indicator of change detection performance.

[0051] F1= = (2) Among them, TP (True Positives) is the number of correctly predicted positive categories, FP (False Positives) is the number of incorrectly predicted positive categories. FN (False Negatives) is the number of incorrectly predicted positive categories.

[0052] Precision (P) refers to the proportion of positive categories predicted by the model to actually be positive categories. The calculation formula is: (3) Recall (R) refers to the proportion of all actual positive categories that are correctly predicted as positive by the model. The calculation formula is: (4) The F1 score is the harmonic mean of precision and recall, which strikes a balance between the two and is often used to deal with unbalanced data sets. When the gap between precision and recall is large, the F1 score will be lower, which forces the model to strike a better balance between precision and recall.

[0053] The comparison results of the present invention and classic or latest methods such as maskRcnn, xbd-baseline, changeos, and Damformer are shown below, as shown in Table 1: Table 1 Comparison results between different methods

[0054] From the comparison results in Table 1 above, it can be seen that by adding the feature mining module 120 and the dynamic label allocation module 130-1, the feature expression ability of the model and the multi-level utilization of the feature: have been significantly improved. Compared with the classic UNET variant network, the overall F1 value of the present invention has increased by about 5.5%. Specifically, in terms of detailed categories, the No-damage (undestroyed) category and the Destroyed (completely destroyed) category have increased most significantly, followed by the Max-damage (large destruction) category. The improvement is also obvious, but it is not difficult to find that the improvement of Min-damage (small destruction) is not particularly obvious compared to other categories. Compared with maskrcnn, the overall F1 has increased by about 3%. Specifically, in terms of a specific category, the improvement trend remains roughly the same.

[0055] Compared with the most recent cutting-edge methods, it can be seen that the method of the present invention is improved compared to changeos and its basic variants, with an overall F1 improvement of about 2.5%. Among them, in different label categories, the performance improvement for completely destroyed labels and large destroyed labels is relatively large, and the performance improvement for undestroyed labels is relatively small. Compared with Damformer, the method of the present invention adopts a faster convergence method, and at the same time improves on the final performance compared to the model.

[0056] In summary, the model of the present invention has advantages over classical methods and recent frontier methods.

[0057] The following shows the ablation experiment results of the feature mining module 120 of the present invention, as shown in Table 2: Table 2 Ablation experiment results of feature mining module 120

[0058] It can be seen from Table 2 (where Base represents the basic model) that the feature mining module 120 mainly improves the information aggregation of the encoder for content representation and position representation in the high-level semantic information of image features by increasing the number of MLP layers. The more layers of the neural network, the easier it is to decode content and position information from high-level semantics, which serves as prior knowledge to guide the network to better approximate the true value, rather than simply approximating a random initial point.

[0059] As the modules are added, the feature mining capability has increased significantly. The comprehensive F1 index and the F1 index of the four categories have increased to varying degrees. The improvement of the max_damage module is the most obvious. As the number of layers increases and the mining capability is enhanced, the F1 score is also further improved. The results are as follows: Figure 4 As shown, the leftmost is the mining module-small, the middle is the mining module-large, and the rightmost is the true value. It can be seen that compared with the true value on the right, the left module has false calls, while the middle module does not have this shortcoming, and the performance has been further improved.

[0060] The following shows the ablation experiment results of the dynamic label allocation module 130-1 of the present invention, as shown in Table 3: Table 3 Ablation experiment results of dynamic label allocation module 130-1

[0061] It can be seen from Table 3 (where Baseline represents the basic model) that the dynamic label allocation module 130-1 plays a vital role in the network training process. The core function of this module is to flexibly adjust the ratio of one-to-many supervision and one-to-one supervision at different stages of training by introducing dynamic adjustment factors. This dynamic adjustment mechanism enables the network to achieve a balance throughout the training process, which can not only enhance the network's representation learning ability, but also effectively accelerate the convergence process. In the early stages of training, the dynamic label allocation module 130-1 tends to increase the ratio of one-to-many supervision, which helps the network better capture the complex patterns and diversity in the data, thereby improving its representation learning ability. In this way, the network can more comprehensively understand the characteristics of the input data and lay a solid foundation for the building detection task. As the training deepens, the dynamic label allocation module 130-1 gradually reduces the ratio of one-to-many supervision and increases the ratio of one-to-one supervision. This adjustment strategy helps the network to focus more on optimizing the prediction results of each sample in the later training, thereby accelerating the convergence process. In this way, the network can significantly shorten the training time while maintaining high detection accuracy, and achieve more efficient end-to-end detection.

[0062] It can also be seen that with the addition of the dynamic adjustment factor, the total value of F1 has increased significantly, and the F1 index of each category has also increased at different levels, which shows that the dynamic adjustment factor has a great effect on improving the model's capabilities. This phenomenon further verifies the effectiveness of the dynamic label allocation module 130-1, which not only improves the performance of the overall model at a macro level, but also optimizes the detection accuracy of each category at a micro level.

[0063] Specifically, the introduction of dynamic adjustment factors enables the model to more flexibly adjust its learning strategy when processing samples of different categories. For those categories that are difficult to distinguish, the dynamic adjustment factor can help the model better capture the characteristics of these categories by increasing the strength of the supervisory signal, thereby improving the detection accuracy. For those categories that are relatively easy to distinguish, the dynamic adjustment factor can reduce unnecessary computational overhead while maintaining high accuracy, further improving the overall efficiency of the model.

[0064] In summary, in view of the complex blending of foreground and background in current buildings, less data distribution, and the problem of missed detection and false detection in change detection, the present invention proposes an environmental change detection system and method based on a priori prompts to dynamically adjust the network. By adding a feature mining module and a dynamic label allocation module, the performance of the model for building change detection is effectively improved. Experiments have proved that this method can reduce the impact of complex background on building change detection, strengthen the extraction of building features, enhance the multi-level utilization of features, and is feasible.

[0065] In response to the problems of target loss and low detection accuracy in the existing environmental change detection and damage assessment networks for urban buildings, the present invention proposes a neural network model that is oriented to complex backgrounds, dynamic adjustment, and deep feature mining. First, a feature mining module is introduced to deeply mine the high-level semantic information of the image for complex backgrounds, decouple the content and location information, and use it as a priori guidance to initialize the query in the decoder to better approximate the true value of the label. Then a dynamic adjustment module is proposed to balance one-to-one supervision and one-to-many supervision to achieve the best performance of the network model. The network proposed by the present invention has greatly improved the accuracy of environmental change detection of buildings in complex backgrounds, while taking into account the convergence speed of the model and reducing redundant calculations in training.

[0066] According to the present invention, there is also provided an electronic device, such as Figure 5 The electronic device 500 may include: a processor 510, a communication interface 520, a memory 530 and a communication bus 540, wherein the processor 510, the communication interface 520 and the memory 530 communicate with each other via the communication bus 540. The processor 510 may call the logic instructions in the memory 530 to execute the steps of the method for detecting environmental changes in the network dynamically adjusting the network based on priori prompts described herein, which will not be described in detail here.

[0067] In addition, according to the present invention, a computer-readable medium, such as a non-transitory computer-readable medium, is also provided for storing the above code or instructions. The computer-readable medium may be a volatile or non-volatile memory or other storage device, which is removably or fixedly connected to a computer, a processor, a controller or other signal processing device to run the code or instructions for performing the operation of the method embodiment or the device embodiment of the present invention.

[0068] The controllers, processors, control circuits, devices, modules, units, multiplexers, logic, interfaces, decoders, drivers, generators, and other signal generation and signal processing features of the embodiments disclosed herein may be implemented, for example, as non-transient logic, which may include hardware, software, or both. When implemented at least in part as hardware, the controllers, processors, control circuits, devices, modules, units, multiplexers, logic, interfaces, decoders, drivers, generators, and other signal generation and signal processing features may be, for example, any of a variety of integrated circuits, including but not limited to application specific integrated circuits, field programmable gate arrays, combinations of logic gates, systems on a chip, microprocessors, or other types of processing or control circuits.

[0069] Although the present teaching has been described with respect to specific embodiments, it is obvious to those skilled in the art that, according to the present disclosure, various changes and modifications may be made without departing from the spirit and scope of the present disclosure as defined in the appended claims. In addition, the embodiments may be combined to form other embodiments.

Claims

1. An environmental change detection system for dynamically adjusting the network based on prior prompts, characterized in that: It includes an encoder module, a feature mining module, a decoder module and a change detection module, among which, The encoder module is configured to: extract and encode features of an input image to obtain fine-grained features; send the fine-grained features as K, V embedding vectors to a decoder module; and send the fine-grained features to a feature mining module; The feature mining module is configured to: perform mining operations on the fine-grained features to obtain position information and content information; send the position information and the content information as prior hints to the decoder module; The decoder module is configured to: assign the prior hint to the query to obtain a new query; in each iteration of training, based on the weight ratio of one-to-many label matching and one-to-one label matching, perform a first operation on the new query and the K, V embedding vector to obtain an output vector and update the new query; when a predetermined condition is met, stop iteration and obtain a final output vector; The change detection module is configured to: perform a second operation on the final output vector and the mask information obtained based on the fine-grained features, and detect changes in the input image based on a result of the second operation, The decoder module further comprises a dynamic label assignment module, and the dynamic label assignment module is configured to change the weight ratio of the one-to-many label matching and the one-to-one label matching during each iteration of training.

2. The environmental change detection system for dynamically adjusting the network based on priori prompts according to claim 1 is characterized in that: As the number of iterations increases, the weight ratio of the one-to-many label matching and the one-to-one label matching continues to decrease.

3. The environmental change detection system for dynamically adjusting the network based on priori prompts according to claim 1 is characterized in that: The predetermined condition is reaching a set maximum number of iterations.

4. The environmental change detection system for dynamically adjusting the network based on priori prompts according to claim 1, characterized in that: Extracting and encoding the input image to obtain the fine-grained features includes: performing a backbone operation on the input image to obtain coarse-grained features; and encoding the coarse-grained features to obtain the fine-grained features.

5. The environmental change detection system for dynamically adjusting the network based on priori prompts according to claim 1 is characterized in that: The mask information is obtained by upsampling the fine-grained features.

6. The environmental change detection system for dynamically adjusting the network based on priori prompts according to claim 1, characterized in that: The first operation is a cross-attention operation.

7. The environmental change detection system for dynamically adjusting the network based on priori prompts according to claim 1 is characterized in that: The second operation is a dot product operation.

8. A detection method for an environment change detection system based on a priori prompts and dynamically adjusting the network according to any one of claims 1 to 7, characterized in that: include: Through the encoder module, feature extraction and encoding are performed on the input image to obtain fine-grained features; Send the fine-grained features as K, V embedding vectors to the decoder module; send the fine-grained features to the feature mining module; The fine-grained features are mined by the feature mining module to obtain position information and content information; the position information and the content information are sent to the decoder module as a priori prompts; Assigning the prior hint to a query through the decoder module to obtain a new query; In each iteration of training, based on the weight ratio of one-to-many label matching and one-to-one label matching, the new query and the K, V embedding vector are subjected to a first operation to obtain an output vector and update the new query; when a predetermined condition is met, the iteration is stopped and a final output vector is obtained; The change detection module performs a second operation on the final output vector and the mask information obtained based on the fine-grained features, and detects changes in the input image based on the result of the second operation, The detection method further comprises: changing the weight ratio of the one-to-many label matching and the one-to-one label matching during each iteration of training through a dynamic label allocation module included in the decoder module.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the program, the steps of the detection method of the environment change detection system for dynamically adjusting the network based on prior prompts as described in claim 8 are implemented.

10. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the detection method of the environmental change detection system for dynamically adjusting the network based on priori prompts as claimed in claim 8 are implemented.

Images