System and method for detecting environmental changes based on dynamic adjustment of network based on prior hints

Abstract

The application provides an environment change detection system and method based on prior prompt dynamic adjustment of a network, wherein an encoder module is configured to: perform feature extraction and encoding on an input image to obtain fine-grained features; take the fine-grained features as K, V embedding vectors; 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 and send the position information and the content information as prior prompts to a decoder module, the decoder module is configured to: give the prior prompts to a query to obtain a new query; perform a first operation on the new query and the K, V embedding vectors based on the weight proportion of one-to-many label matching and one-to-one label matching to obtain an output vector and update the new query, and a change detection module is configured to: perform a second operation on the last output vector and mask information obtained based on the fine-grained features, and detect changes in the input image based on the result of the second operation.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer, in particular, to an environment change detection system and method based on prior prompt dynamic adjustment network. BACKGROUND

[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 necessary. Remote sensing technology quickly captures the physical state changes of buildings and infrastructure by comparing and analyzing images of the same area at different times, providing key decision support for rescue teams, promoting the precise allocation of rescue resources, and providing important planning basis for post-disaster reconstruction. Satellite remote sensing plays an important role in disaster monitoring due to its regular coverage, wide-area monitoring and high-resolution imaging. For example, LandSat series, Ikonos, QuickBird, WorldView satellites 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, which can quickly determine the changes of buildings by comparing and analyzing satellite images at different times, and capture the damage caused by disasters to buildings. This technology not only helps to assess the impact of disasters in a timely manner, but also provides important basis for subsequent rescue operations. The results of change detection provide key data support for rescue decision-making, enabling rescue resources to be accurately allocated to where they are most needed, and providing important planning basis for post-disaster reconstruction. Building change detection, as a type of change detection, is important because it can quickly capture the damage to buildings, provide key decision support for rescue teams, and help post-disaster reconstruction. By identifying which buildings can be repaired and which need to be demolished and rebuilt, the efficiency of reconstruction is improved. Deep learning technology is used to analyze remote sensing images, automatically learn and extract complex features, improve the accuracy and efficiency of change detection, and thus exhibit 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 differencing, image ratio, image regression, and change vector analysis (CVA) have achieved certain results in some application scenarios, they usually ignore the local information of images, resulting in high sensitivity to noise and registration errors. In addition, these methods rely on manually designed fixed features, which 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 more rich and robust image representations through their powerful feature extraction capabilities, thus showing significant advantages in change detection or damage assessment tasks. Deep learning models can identify 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 of change features directly from raw pixel data without complex feature engineering, which further simplifies the process and improves performance. However, current deep learning methods are not effective when further requirements are required, such as in the process of building change detection, current network models cannot balance the enhancement of model representation ability while accelerating convergence. SUMMARY

[0004] To solve the above problems, the present application provides a building change detection system and method based on prior prompt dynamic adjustment of network.

[0005] According to an aspect of the present application, there is provided an environment change detection system for dynamically adjusting a network based on prior hints, comprising an encoder module, a feature mining module, a decoder module and a change detection module, wherein the encoder module is configured to: perform feature extraction and encoding on an 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 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 hints to a query to obtain a new query; at each iteration of training, perform a first operation on the new query and the K, V embedding vectors based on a weight ratio of one-to-many label matching and one-to-one label matching to obtain an output vector and update the new query; stop iteration and obtain a final output vector when a predetermined condition is met; the change detection module is configured to: perform a second operation on the final output vector and mask information obtained based on the fine-grained features, and detect changes in the input image based on a result of the second operation, wherein the decoder module further comprises 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 at each iteration of training.

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

[0007] Further, the predetermined condition is to reach a set maximum number of iterations.

[0008] Further, the feature extraction and encoding on the input image to obtain fine-grained features comprises: 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.

[0009] Further, 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 an aspect of the present application, there is provided a detection method of an environment change detection system based on dynamic adjustment of a network, comprising: performing feature extraction and encoding on an input image by an encoder module to obtain fine-grained features; sending the fine-grained features to a decoder module as K, V embedding vectors; sending the fine-grained features to a feature mining module; mining the fine-grained features by the feature mining module to obtain position information and content information; sending the position information and the content information to the decoder module as prior hints; obtaining a new query by assigning the prior hints to a query by the decoder module; performing a first operation on the new query and the K, V embedding vectors based on a weight ratio of one-to-many label matching and one-to-one label matching to obtain an output vector and update the new query at each iteration of training; stopping iteration and obtaining a final output vector when a predetermined condition is met; performing a second operation on the final output vector and mask information obtained based on the fine-grained features by a change detection module, and detecting changes of the input image based on a result of the second operation, wherein the detection method further comprises: changing the weight ratio of the one-to-many label matching and the one-to-one label matching at each iteration of training by a dynamic label assignment module included in the decoder module.

[0013] According to an aspect of the present application, there is provided an electronic device comprising 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 detection method of the environment change detection system based on dynamic adjustment of a network as above when executing the program.

[0014] According to an aspect of the present application, there is provided a non-transitory computer-readable storage medium having a computer program stored thereon, wherein the computer program is executable by a processor to implement the steps of the detection method of the environment change detection system based on dynamic adjustment of a network as above.

[0015] According to the present application, a dynamic network adjustment environment change detection system and method based on prior prompt are provided, wherein compared with random initialization, a method of reinitializing query is proposed, a prompt is added to the query, prior knowledge is given to the query, and the query contains content and location information. In order to better serve the query learning, the location and content are as accurate as possible, in order to obtain finer granularity of location and content, a classifier is added after the encoder, and in order to better improve the accuracy of the classifier, a feature mining module is added before the classical classifier. This effectively solves the false detection problem that easily occurs in building change detection under remote sensing images; in addition, a dynamic candidate box number adjustment strategy is proposed, a function with epoch as the independent variable is constructed, a one-to-many label assignment strategy and a one-to-one label assignment strategy are controlled through the training round number, one-to-many label matching supervises one-to-one label matching, one-to-many label matching can better enhance the learning representation ability of the model, one-to-one label matching can accelerate the convergence of the model, realizes end-to-end detection, and simplifies the training. Through dynamic function adjustment, the problem of low feature learning efficiency of only one-to-one supervision mechanism is solved due to the difference in the number of buildings in different images. BRIEF DESCRIPTION OF DRAWINGS

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

[0017] Figure 2 is a schematic diagram of an environment change detection system based on prior prompt dynamic network adjustment according to an embodiment of the present application.

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

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

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

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

[0022] The present application mainly aims to solve the problem of low building change detection performance in the current remote sensing field. In the photographed satellite image, buildings often show discrete distribution characteristics, and the complex background and foreground are complex. There are significant missing and false calls for building foreground targets. To solve this problem, a neural network model based on dynamic adjustment of feature depth mining is proposed. Through dynamic adjustment factor, the depth mining module improves the performance of the neural network. The present application aims to enhance the model representation capability, proposes a deep feature mining module and a dynamic label assignment module, and further solves the problem of low accuracy and recall rate of the previous model in the remote sensing data set for building change detection, and improves the performance of target detection.

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

[0024] Referring to Figure 1 , Figure 1 The upper left corner of is a scene graph before the disaster occurs (model input), Figure 1 The lower right corner of is a scene graph after the disaster occurs (model output), and the input picture is processed (for example, a backbone network) to obtain coarse extraction features (coarse-grained features). By processing the coarse extraction features, fine extraction features (fine-grained features) are obtained. The operation of fine-grained features can include three kinds: first, fine-grained features are used for subsequent feature miners for feature mining, so as to obtain content information (content) and location information (location) for operation in the subsequent decoder (as prior information, the query in the decoder is given to approximate the objective distribution of the real world true value); second, fine-grained features are used as K, V embedding vectors for operation in the subsequent decoder; third, fine-grained features are operated to obtain mask information (mask) for subsequent mask calculation. In the present model, a dynamic adjustment factor μ is introduced to control the weight proportion of one-to-many label matching and one-to-one label matching, to ensure the learning ability of the model while taking into account accelerating the convergence of the model.

[0025] Figure 2 is a schematic diagram of an environment change detection system 100 based on prior prompt dynamic adjustment network according to an embodiment of the present application.

[0026] Referring to 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 embodiments, the encoder module 110 can be configured to perform feature extraction and encoding on the input image to obtain fine-grained features; send the fine-grained features to the decoder module 130 as K, V embedding vectors; and send the fine-grained features to the feature mining module 120.

[0028] In embodiments, the feature mining module 120 can be configured to perform mining operations on the fine-grained features to obtain location information and content information; and send the location information and the content information to the decoder module 130 as prior hints.

[0029] According to the present application, by the operation of the feature mining module 120, the fine-grained features are subjected to mining operations to obtain location information and content information; and the location information and the content information are sent to the decoder module as prior hints. By adding the location information and the content information as prior knowledge to the query in the processing flow, unlike the most traditional random initialization, this helps to further match the true values of the dataset labels, accelerating model training and convergence.

[0030] In embodiments, the decoder module 130 can be configured to give the prior hints to the query to obtain a new query; in each iteration of training, based on a weight ratio (which can be regarded as a dynamic adjustment factor, which will be described in detail below) 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 vectors to obtain an output vector and update the new query; and when a predetermined condition is met, stop iteration and obtain a final output vector.

[0031] In embodiments, the change detection module 140 can be configured to perform a second operation on the final output vector of the decoder module 130 and 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 application, the predetermined condition can be that the iterative training reaches a set maximum number of iterations, but the present application is not limited thereto, for example, the predetermined condition can be set as detecting a change performance (for example, F1 mentioned below) to reach a predetermined threshold.

[0033] According to the present application, performing feature extraction and encoding on the input image to obtain fine-grained features can include performing a backbone operation on the input image to obtain coarse-grained features; and encoding the coarse-grained features (for example, through an L-layer encoder) to obtain the fine-grained features.

[0034] According to the present application, the mask information can be obtained by up-sampling and restoring the fine-grained features.

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

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

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

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

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

[0040] In the present application, one-to-many label matching and one-to-one label matching are introduced, and the one-to-one paradigm is supervised by the one-to-many paradigm, so that the model has both feature expression capability and model fast convergence capability. At this time, the input in the decoder is a query that obtains prior knowledge, which can be called a new query. The new query and the fine-grained feature (K, V embedding vector as above) are subjected to attention operation.

[0041] Among them, one-to-many label matching refers to the matching of multiple predicted frames with one real frame in the neural network, and the most suitable candidate frame is finally selected by the method of non-maximum suppression. One-to-one label matching refers to the matching of only one predicted frame with one real frame in the neural network.

[0042] The advantage of the one-to-many label matching paradigm is that the learning representation capability of the model is strong, and the disadvantages are that a large number of anchor frames need to be designed manually in the early stage of the model, the parameters are redundant, and are 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 more concise, and the model is easy to converge. The disadvantage is that only one candidate frame is used for judgment, the expression capability of the model is reduced, and confusion phenomenon is easy to occur in the reasoning process.

[0043] The present application 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 training process of the model, in order to maintain the feature expression learning capability of the model, the one-to-many way is used for supervision. With the increase of the number of training rounds, the weight proportion of one-to-many label matching and one-to-one label matching is changed by a dynamic adjustment factor, so as to ensure the learning capability of the model while taking into account accelerating the convergence of the model; in the later stage of model training, the learning capability of the neural network model is relatively saturated, the dynamic adjustment factor is changed, at this time the model tends to the traditional one-to-one label matching, which can accelerate the convergence of the neural network, and at the same time make the training process more concise.

[0044] Among them, the calculation of the dynamic adjustment factor is as follows:

[0045] (1)

[0046] In the above formula (1), T is the total number of training rounds of the model, t is the current training round, and n is the initial weight ratio of one-to-many supervision and one-to-one supervision. Through the above dynamic adjustment factor The balanced label assignment strategy is realized. That is, through the above equation, the dynamic adjustment factor The initial value of n is constantly decreasing as the iteration proceeds.

[0047] Therefore, the dynamic label assignment module 130-1 of the decoder module 130 can be configured to change the weight ratio of one-to-many label matching and one-to-one label matching at each iteration of training.

[0048] According to the above concept of dynamic balanced label assignment, the following shows the algorithm of dynamic balanced label assignment:

[0049]

[0050] In order to verify the effect achieved by the technical scheme of the present application, the present application uses the XBD dataset for testing.

[0051] The XBD dataset is an innovative large-scale satellite image dataset designed to evaluate building damage after natural disasters. The dataset contains 22,068 satellite images with 1024x1024 pixel resolution from 19 different natural disaster events, covering a geographical area of over 45,000 square kilometers. The image data is collected from WorldView-3 satellites, with a spatial resolution of 0.3 meters / pixel. The XBD dataset introduces a four-level damage annotation scale, from no damage (0) to complete destruction (4), providing a standardized evaluation tool for building damage assessment.

[0052] This test uses 2*4090 24G and 1 A100 40G to complete, for the convenience of testing, the dataset is converted to COCO better via multi-model detection. The model is based on the PyTorch framework and trained in the cloud. In the training process, the Stochastic Gradient Descent (SGD) optimizer is used, the initial learning rate is 1e-2, and the learning rate is decayed to 2e-3 as the training progresses. The total training rounds of the training process are 500 (epochs).

[0053] This test uses F1 as the evaluation index of change detection performance.

[0054] F1= = (2)​

[0055] where TP (True Positives) is the number of correctly predicted positive classes, FP (False Positives) is the number of incorrectly predicted positive classes, and FN (False Negatives) is the number of incorrectly not predicted positive classes.

[0056] Precision (P) is the proportion of positive classes that are actually positive in the positive classes predicted by the model, and the calculation formula is:

[0057] (3)

[0058] Recall (R) is the proportion of all actual positive classes that are correctly predicted as positive classes by the model, and the calculation formula is:

[0059] (4)

[0060] F1 score is the harmonic mean of precision and recall, which balances between the two. When the precision and recall are significantly different, the F1 score will be relatively low, which forces the model to achieve a better balance between precision and recall.

[0061] The following shows the comparison results of the present application and maskRcnn, xbd-baseline, changeos, Damformer and other classic or latest methods, as shown in Table 1:

[0062] Table 1 Comparison results between different methods

[0063]

[0064] From the comparison results of Table 1 above, it can be seen that after adding the feature mining module 120 and the dynamic label assignment module 130-1, the feature expression ability of the model and the multi-level utilization of the features have been significantly improved. Compared with the classic UNET variant network, the overall F1 value of the present application is improved by about 5.5%, and in detail, the No-damage (undamaged) class and the Destroyed (completely destroyed) class are improved the most, followed by the Max-damage (larger damage) which is also improved significantly, but for the Min-damage (smaller damage), it is not difficult to find that compared with other classes, the improvement is not particularly obvious. Compared with maskrcnn, the overall F1 is improved by about 3%, and in detail, the improvement trend of a specific class is roughly consistent.

[0065] Compared with the recent frontier method, it can be seen that the method of the present application is improved compared with changeos and its basis variants, and the F1 overall is improved by about 2.5%. Among them, in different label categories, the performance improvement for completely destroyed labels and larger destroyed labels is relatively large, and the performance improvement for undamaged labels is relatively small. Compared with Damformer, the method of the present application adopts a faster convergence method, and at the same time, the final performance is improved compared with the model.

[0066] In summary, the model of the present application has more advantages than the classical method and the recent frontier method.

[0067] The following shows the ablation experiment results of the feature mining module 120 of the present application, as shown in Table 2:

[0068] Table 2 Ablation experiment results of the feature mining module 120

[0069]

[0070] As shown in Table 2 (where Base represents the base model), the feature mining module 120 mainly improves the information aggregation of the encoder for the content representation and position representation in the high-level semantic information of the image features by increasing the number of layers of the MLP. The more layers of the neural network, the easier it is to decode the content and position information from the high-level semantics as prior knowledge to guide the network to better approximate the true value, rather than simply approximating the true value from a random initial point.

[0071] With the addition of the module, it can be found that the mining ability of the feature has increased significantly, and the comprehensive F1 index and the F1 index of the four categories have all increased to different degrees. Among them, the max_damage module has the most obvious improvement. With the increase of the number of layers and the enhancement of the mining ability, the F1 score is further improved, as shown in the results of Figure 4 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 alarm, and the middle module does not have this shortcoming, and the performance is further improved.

[0072] The following shows the ablation experiment results of the dynamic label assignment module 130-1 of the present application, as shown in Table 3:

[0073] Table 3 Ablation experiment results of the dynamic label assignment module 130-1

[0074]

[0075] As can be seen from Table 3 (where Baseline represents the base model), the dynamic label assignment module 130-1 plays a crucial 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 a dynamic adjustment factor. This dynamic adjustment mechanism enables the network to achieve a balance throughout the training process, enhancing both the network's representation learning ability and effectively accelerating the convergence process. In the early stages of training, the dynamic label assignment module 130-1 tends to increase the proportion of one-to-many supervision, which helps the network better capture complex patterns and diversity in the data, thereby improving its representation learning ability. In this way, the network can more comprehensively understand the features of the input data, laying a solid foundation for the building detection task. As the training progresses, the dynamic label assignment module 130-1 gradually reduces the proportion of one-to-many supervision and increases the proportion of one-to-one supervision. This adjustment strategy helps the network 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, achieving more efficient end-to-end detection.

[0076] At the same time, it can also be seen that with the addition of the dynamic adjustment factor, the total value of F1 has significantly improved, and the F1 indicators of each class have also improved at different levels, indicating that the dynamic adjustment factor has greatly improved the model's ability. This phenomenon further verifies the effectiveness of the dynamic label assignment module 130-1, which not only improves the overall performance of the model macroscopically but also optimizes the detection accuracy of each class microscopically.

[0077] Specifically, the introduction of the dynamic adjustment factor enables the model to more flexibly adjust its learning strategy when processing samples of different classes. For classes that are difficult to distinguish, the dynamic adjustment factor can help the model better capture the features of these classes by increasing the intensity of the supervision signal, thereby improving the detection accuracy. For relatively easy-to-distinguish classes, the dynamic adjustment factor can reduce unnecessary computational overhead while maintaining high accuracy, further improving the overall efficiency of the model.

[0078] In summary, in view of the current problem of complex foreground and background blending in buildings, less data distribution, and missed detection and false detection in change detection, the present application proposes a change detection system and method for environment change based on prior prompt dynamic adjustment network. By increasing the feature mining module and the dynamic label assignment module, the performance of the model for building change detection is effectively improved. Experiments show that this method can reduce the influence of complex background on building change detection, strengthen the extraction of building features, enhance the multi-level use of features, and has feasibility.

[0079] The present application aims at the problems of target loss and low detection accuracy existing in the current environment change detection and damage assessment network for urban buildings, and proposes a neural network model facing complex background, dynamic adjustment and feature depth mining. First, a feature mining module is introduced, which deeply mines the high-level semantic information of the image in the complex background, decouples the content and position information, and uses them as priori guidance to initialize the query in the decoder, so as to better approximate the label true value. Then, a dynamic adjustment module is proposed to balance one-to-one supervision and one-to-many supervision, so that the network model performance reaches the best. The network proposed in the present application greatly improves the accuracy of building environment change detection in complex background, and also considers the convergence speed of the model and reduces the redundant calculation of training.

[0080] According to the present application, an electronic device is also provided, such as Figure 5 The electronic device 500 can 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 can communicate with each other through the communication bus 540. The processor 510 can call the logical instructions in the memory 530 to execute the steps of the environment change detection method based on the prior prompt dynamic adjustment network described herein, which will not be described in detail here.

[0081] In addition, according to the present application, a computer readable medium, such as a non-transitory computer readable medium, is also provided for storing the above-mentioned codes or instructions. The computer readable medium can be a volatile or non-volatile memory or other storage device, which is removably or fixedly coupled to a computer, a processor, a controller or other signal processing device, to run the codes or instructions for performing the operations of the method embodiments or device embodiments herein.

[0082] The controllers, processors, control circuits, devices, modules, units, multiplexers, logics, interfaces, decoders, drivers, generators and other signal generation and signal processing features of the embodiments disclosed herein can be implemented, for example, as non-transient logic, which can include hardware, software or both. When implemented at least partially as hardware, the controllers, processors, control circuits, devices, modules, units, multiplexers, logics, interfaces, decoders, drivers, generators and other signal generation and signal processing features can be, for example, any of various integrated circuits, including but not limited to application specific integrated circuits, field programmable gate arrays, combinations of logic gates, system on chips, microprocessors or other types of processing or control circuits.

[0083] While the present teachings have been described in conjunction with specific embodiments thereof, it is evident that many alternatives, modifications and variations will be apparent to those skilled in the art in light of the present disclosure and the foregoing description is intended to embrace all such alternatives, modifications and variations as fall within the scope of the present teachings as defined by the appended claims. Furthermore, the embodiments can be combined to form other embodiments.

Claims

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

2. The system according to claim 1, wherein With the increase of the number of iterations, the weight ratio of the one-to-many label matching and the one-to-one label matching decreases constantly. 3.The environment change detection system based on prior hint dynamic adjustment network of claim 1, wherein, The predetermined condition is to reach a set maximum number of iterations.

4. The system of claim 1, wherein the system is configured to: Performing feature extraction and encoding on the input image to obtain the fine-grained features comprises: 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 system of claim 1, wherein the system is configured to: The mask information is obtained by upsampling the fine-grained features.

6. The system of claim 1, wherein the system is configured to: 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, characterized in that: The second operation is a dot product operation.

8. A detection method of a change-in-environment detection system based on a priori hint dynamic adjustment network according to any one of claims 1 to 7, characterized by, The system comprises: performing feature extraction and encoding on an input image by the encoder module to obtain fine-grained features; sending the fine-grained features as K, V embedding vectors to the decoder module; and sending the fine-grained features to the feature mining module; performing mining on the fine-grained features by the feature mining module to obtain position information and content information; and sending the position information and the content information as prior hints to the decoder module; assigning the prior hints to a query by the decoder module to obtain a new query; in each iteration of training, performing a first operation on the new query and the K, V embedding vectors based on a weight ratio of one-to-many label matching and one-to-one label matching to obtain an output vector and update the new query; and stopping iteration and obtaining a final output vector when a predetermined condition is met; performing a second operation on the final output vector and mask information obtained based on the fine-grained features by the change detection module, and detecting changes in the input image based on a result of the second operation, The detection method further comprises: changing, by a dynamic label assignment module included in the decoder module, a weight ratio of the one-to-many label matching and the one-to-one label matching at each iteration of the training.

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, The processor implements the steps of the detection method of the environment change detection system based on the prior prompt dynamic adjustment network according to the program.

10. A non-transitory computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to implement the steps of the detection method of the environment change detection system based on the prior prompt dynamic adjustment network.

Images