CN-121982446-A - Point supervision change detection model training method based on active learning

CN121982446ACN 121982446 ACN121982446 ACN 121982446ACN-121982446-A

Abstract

The invention discloses a point supervision change detection model training method, device, medium and equipment based on active learning, and aims to solve the model training problems caused by low point information labeling efficiency and sparse point labeling dispersion in point supervision change detection. The method is a man-machine interaction type multi-iteration training method, and a model is trained based on an integral loss function by adding an auxiliary decoder and a pseudo tag generation module to the model to be trained. And taking the sampling, labeling, training and predicting as one round, screening high-information-content pixel point labeling by a sampling algorithm, expanding a label set, enabling the initial value of a pseudo label to be consistent with the labeling label, and monitoring an original decoder and an auxiliary decoder by the labeling label and the pseudo label respectively after each preset round of updating. After multiple iterations, training a high-performance model by using a small number of point marks is realized. The method greatly reduces the labeling cost, improves the detection precision and robustness of the model to the change region, and is suitable for the double-time-phase remote sensing image change detection scene.

Inventors

LIU GANCHAO
HUANG JINCHENG
Xu Yetong
ZHAO YANG
YUAN YUAN

Assignees

西北工业大学

Dates

Publication Date: 20260505
Application Date: 20251229

Claims (10)

1. The point supervision change detection model training method based on active learning is characterized by comprising the following steps of: Adding an auxiliary decoder and a pseudo tag generation module to the change detection model to be trained to obtain a point supervision change detection model based on active learning; Training an auxiliary decoder and an original decoder of a change detection model to be trained based on a consistency loss function, and training a point supervision change detection model based on active learning based on an integral loss function; In the training process of the point supervision change detection model based on active learning in the integral loss function training, the method comprises the following steps: Collecting a preset number of pixel points on each preprocessed double-phase remote sensing image through a random sampling/sampling algorithm, and marking the pixel points obtained by sampling to obtain marked pixel points; adding the marked pixel points into a marked point label set to obtain an initialized pseudo label, wherein the initial value of the pseudo label is consistent with the label value of the marked point label set; Monitoring the training of the original decoder by adopting the marked point label set, and monitoring the training of the auxiliary decoder by adopting a pseudo label, wherein the pseudo label comprises an initialization pseudo label which is used for the first time and one group of pseudo labels which are updated by the pseudo label generating module after each preset round in the training process, so as to obtain a trained change detection model; extracting output of an original decoder of the trained change detection model as a prediction result, inputting the prediction result into the sampling algorithm to serve as a new pixel point, obtaining a new pseudo tag and a new marked pixel point based on the new pixel point, and respectively executing subsequent supervised training on the original decoder and the auxiliary decoder based on the new pseudo tag and the new marked pixel point until the preset number of active learning training rounds is reached, so as to obtain a final point supervised change detection model based on active learning.
2. The method for training a point supervised variation detection model based on active learning as set forth in claim 1, wherein the expression of the consistency loss function is: Where n represents the number of pixels, Representing the output of the original decoder to pixel i, Representing the output of the auxiliary decoder to pixel i.
3. The active learning-based point supervision change detection model training method of claim 2, further comprising: Training the auxiliary decoder using lovas loss functions; wherein, lovas loss function has the expression: wherein n e denotes the number of labeled pixels, Represents the extended Jaccard loss, and m (i) represents the range loss for pixel i.
4. The active learning based point supervised variation detection model training method as set forth in claim 1, wherein the label generation module is determined based on a region growing algorithm comprising: Calculating the prediction confidence of the pixel, wherein the calculation expression of the prediction confidence of the pixel is as follows: Where p (i) denotes the output of the model to pixel i; generating a pseudo tag based on preset region growing conditions, wherein the region growing conditions are as follows: Wherein E (l) represents the marked pixel l, namely the category l of the growing seeds, E (u) represents the pixel u in the neighborhood of the pixel l, namely the category of the pixel to be grown, The prediction result of the model on the change class of the pixel u is represented, confidence (p (u)) represents the prediction confidence of the pixel u, and τ represents the confidence threshold.
5. The method for training the active learning-based point supervised variation detection model as set forth in claim 1, wherein the sampling algorithm includes: removing marked pixels in each pair of double-phase remote sensing images to obtain a new double-phase remote sensing image; scoring all pixel points in the new double-phase remote sensing image according to a preset standard scoring formula; Sorting from high to low according to scores, and randomly sampling a specified number of pixel points from pixels with a certain proportion of the scores at the front; The preset standard scoring formula is as follows: where p (i) denotes the output of the model to pixel i.
6. The method of claim 1, wherein the change detection model is SAM-CD, and the network structure includes an encoder of the FastSAM model with frozen weights, an adapter module of the encoder, an original decoder branch, and an auxiliary decoder branch: Wherein the auxiliary decoder has the same structure as the original decoder.
7. The active learning based point supervised variation detection model training method as recited in claim 1, wherein the overall loss function includes a summed cross entropy loss function, and the lovas loss function and the consistency loss function as recited in claim 3.
8. The utility model provides a point supervision change detection model trainer based on initiative study which characterized in that includes: the model construction module is used for adding an auxiliary decoder and a pseudo tag generation module for the change detection model to be trained to obtain a point supervision change detection model based on active learning; The model training module is used for training the auxiliary decoder and the original decoder of the change detection model to be trained based on the consistency loss function, and training the point supervision change detection model based on active learning based on the integral loss function; In the training process of the point supervision change detection model based on active learning in the integral loss function training, the method comprises the following steps: Collecting a preset number of pixel points on each preprocessed double-phase remote sensing image through a random sampling/sampling algorithm, and marking the pixel points obtained by sampling to obtain marked pixel points; adding the marked pixel points into a marked point label set to obtain an initialized pseudo label, wherein the initial value of the pseudo label is consistent with the label value of the marked point label set; Monitoring the training of the original decoder by adopting the marked point label set, and monitoring the training of the auxiliary decoder by adopting a pseudo label, wherein the pseudo label comprises an initialization pseudo label which is used for the first time and one group of pseudo labels which are updated by the pseudo label generating module after each preset round in the training process, so as to obtain a trained change detection model; extracting output of an original decoder of the trained change detection model as a prediction result, inputting the prediction result into the sampling algorithm to serve as a new pixel point, obtaining a new pseudo tag and a new marked pixel point based on the new pixel point, and respectively executing subsequent supervised training on the original decoder and the auxiliary decoder based on the new pseudo tag and the new marked pixel point until the preset number of active learning training rounds is reached, so as to obtain a final point supervised change detection model based on active learning.
9. A computer readable storage medium comprising instructions that when run on a computer cause the computer to perform the active learning based point supervised variation detection model training method of any of claims 1-7.
10. An electronic device, the electronic device comprising: at least one processor, memory, and input output unit; the memory is used for storing a computer program, and the processor is used for calling the computer program stored in the memory to execute the point supervision change detection model training method based on active learning according to any one of claims 1-7.

Description

Point supervision change detection model training method based on active learning Technical Field The application relates to the technical field of remote sensing images, in particular to a point supervision change detection model training method, device, medium and equipment based on active learning. Background The change detection is one of the most important tasks in the intelligent interpretation of the remote sensing image, and the technology can be widely applied to the fields of land coverage, disaster detection, urban construction, military application and the like. The method identifies the difference of objects or phenomena through observation in a period of time in the same geographic area, namely the change between the detection of double-phase images (shot images of the same region at different times). With the development of remote sensing technology, intelligent automatic identification of changes in features is becoming increasingly important for detecting dynamic changes in the earth. The current mainstream change detection method is to construct a change detection model based on a deep learning technology, and then use a pixel-level change label to supervise the training model so as to achieve the performance of actual demands. However, the labeling of pixel-level change labels is costly, limiting the application of the change detection model. For this reason, researchers have proposed a weakly supervised change detection method. Weakly supervised variation detection aims at training a model using low-cost, but not exact, annotation forms (e.g., image-level annotations, target-level annotations, etc.) to achieve performance as close to that of a fully supervised model as possible. According to different labeling forms, the existing weakly supervised change detection methods can be roughly divided into three types, namely a method using image-level labeling, a method using target-level labeling and a method using point-level labeling. The method for detecting the change of the image level annotation only depends on the overall level annotation of the image (namely, whether the change exists in the double-time-phase remote sensing image) to train the model. The labeling form does not need to label the specific change area at the pixel level, so that the labeling cost is low, the efficiency is high, and the labeling method is an important mode for solving the problem of difficulty in labeling a large-scale data set. Mainstream methods typically include locating the change region through an attention mechanism, generating pseudo tag optimization training, and end-to-end optimization based on a classification network, among others. However, the labeling information used by the method is too coarse, the information quantity is small, the model is easy to locate a change area inaccurately, the model is easy to be influenced by background interference or noise, and the detection precision is difficult to reach the actual requirement. A weakly supervised variation detection method using target frame level labeling uses target detection frame level labeling (e.g., a rough bounding box marking the variation region) to train the model. This form of labeling provides finer information than image-level labeling, but still has lower labeling costs than pixel-level labeling, which can help the model learn the location of the change region more accurately. The method has the advantages that good balance is achieved between accuracy and labeling efficiency, interference of background noise can be reduced, and the positioning capability of the model on the change is improved. The mainstream method comprises the steps of generating a pseudo pixel label of a change area through a target frame label, acquiring finer granularity characteristics through combination of instance segmentation, or directly utilizing a target detection frame to optimize network weight. At present, the main defects of the method are that the target frame mark is still insufficient to provide accurate change contour information, the boundary detection capability of a model is possibly limited, and the quality and the size of the frame mark have strong dependence on the stability of a final detection result. A method of point level labeling is used. Weak supervisory change detection (point supervisory change detection) methods using point level labeling rely on labeling a small number of points (e.g., key points within a change region) to train a model. The labeling form provides sparse and accurate change information, the labeling cost is far lower than that of pixel-level labeling, and meanwhile, model optimization can be effectively guided, and a change area is positioned. The method has the advantages that the balance between the labeling quantity and the accuracy is realized, the labeling complexity is reduced, and the model performance under the weak supervision condition is improved. The mainstream method generally generates pseud