CN-121999299-A - Unsupervised PolSAR classification method based on contrast learning

CN121999299ACN 121999299 ACN121999299 ACN 121999299ACN-121999299-A

Abstract

The invention relates to the technical field of image processing, in particular to an unsupervised PolSAR classification method based on contrast learning, which comprises the steps of obtaining original PolSAR image data to form multi-view data, carrying out super-pixel segmentation on an image to generate multi-view element samples corresponding to super-pixels one by one, obtaining potential feature representation of each view, generating a self-representation coefficient matrix of each view through a forward self-representation module, applying cross-view relationship consistency constraint through contrast loss, optimizing the self-representation coefficient matrix of each view, fusing the optimized self-representation coefficient matrix of each view to generate a unified affinity matrix, carrying out spectral clustering to obtain clustering labels of the element samples, and mapping the clustering labels back to pixel levels according to the mapping relation between the element samples and pixels established by the super-pixel segmentation to obtain an image classification result, wherein the accuracy and the robustness of classification are remarkably improved.

Inventors

ZUO XIANYU
LIU QILONG
WANG YADI
Dang Lanxue
XIE YI
QIAO BAOJUN

Assignees

河南大学

Dates

Publication Date: 20260508
Application Date: 20260316

Claims (8)

1. An unsupervised PolSAR classification method based on contrast learning is characterized by comprising the following steps: Acquiring original PolSAR image data, and extracting at least two pixel-level features to form multi-view data; performing superpixel segmentation on the image, and respectively aggregating the characteristics of all pixels in each superpixel under each view to generate a multi-view element sample corresponding to the superpixels one by one; Inputting the multi-view element samples into a corresponding depth feature extraction network to obtain potential feature representation of each view, and generating a self-representation coefficient matrix of each view through a forward self-representation module, wherein each column corresponds to a self-representation coefficient vector of one element sample, and represents the weight of the self-representation coefficient vector linearly reconstructed by all element samples; Based on a relationship consistency contrast learning mechanism, constructing self-expression coefficient vectors of the same element sample under different views into positive sample pairs, applying cross-view relationship consistency constraint through contrast loss, and optimizing a self-expression coefficient matrix of each view; generating a unified affinity matrix by fusing the optimized self-expression coefficient matrix of each view, and performing spectral clustering to obtain a clustering label of the meta-sample; And mapping the clustering label back to the pixel level according to the mapping relation between the meta-sample and the pixel established by the super-pixel segmentation, and obtaining an image classification result.
2. The unsupervised PolSAR classification method according to claim 1, wherein the multi-view data comprises a polarization feature view and a texture feature view.
3. The method for unsupervised PolSAR classification based on contrast learning of claim 1, wherein the aggregation uses a truncated mean aggregation function.
4. The method for unsupervised PolSAR classification based on contrast learning of claim 1, wherein obtaining the self-expression coefficient matrix comprises: carrying out L2 normalization on potential features of all element samples under each view to obtain a normalized feature matrix; calculating an inner product between any two element sample features in the normalized feature matrix to generate an initial similarity matrix; an adaptive soft threshold operator controlled by a learnable parameter is applied to the initial similarity matrix, and a self-representative coefficient matrix is output in forward propagation.
5. The method for unsupervised PolSAR classification based on contrast learning according to claim 1, wherein said optimizing the self-expression coefficient matrix of each view comprises: The relation consistency contrast learning mechanism adopts InfoNCE loss functions, wherein the self-expression coefficient vector of the same meta-sample under different views is used as a positive sample pair, and the self-expression coefficient vector of the meta-sample and other meta-samples under any view is used as a negative sample pair; the self-representing structures of the views are driven to align in the shared subspace by maximizing the similarity between positive pairs of samples and minimizing the similarity between negative pairs of samples, thereby optimizing the self-representing coefficient matrix of the views.
6. The method for unsupervised PolSAR classification based on contrast learning according to claim 1, wherein the generating a unified affinity matrix from the self-expression coefficient matrix of each view after fusion optimization comprises: Based on the self-expression coefficient matrix of each view, weighting and fusing each view through the learnable attention weight; the attention weight is dynamically generated by a lightweight neural network according to the statistical characteristics or structural information of the self-expression coefficient matrix of each view and is used for representing the contribution degree of the view to the final clustering result; and summing the weighted view self-expression coefficient matrixes to obtain the unified affinity matrix.
7. The method of claim 1, wherein the depth feature extraction network is a multi-view self-encoder structure and the training process of the network includes minimizing reconstruction losses between each view input and output.
8. The method of unsupervised PolSAR classification based on contrast learning of claim 1, further comprising, after obtaining the self-representation coefficient matrix and the latent feature representation: Regularization constraints including elastic mesh regularization and graph regularization are applied to the self-representative coefficient matrix and the latent feature representation, respectively.

Description

Unsupervised PolSAR classification method based on contrast learning Technical Field The invention relates to the technical field of image processing, in particular to an unsupervised PolSAR classification method based on contrast learning. Background The polarized synthetic aperture radar (PolSAR) has all-weather and all-day imaging capability, and image classification is a key link of remote sensing interpretation. The existing supervised classification method based on deep learning is high in precision, but depends on a large number of manual labeling samples, and the PolSAR data labeling cost is high, the specialization is strong, the acquisition is difficult, and the practical application is limited. Therefore, an unsupervised classification method without labeling becomes an important research direction. The depth multi-view subspace clustering (DMVSC) learns the shared subspace structure by fusing multi-source features such as polarization, texture and the like in the PolSAR image, and shows potential in unsupervised classification. However, when applied to large-scale high-resolution PolSAR images, the following key technical bottlenecks are still faced: (1) The expandability is insufficient, namely, a global self-expression matrix is directly constructed at the pixel level, spectral clustering is carried out, the calculation and memory cost is rapidly increased along with the image scale, and the method is difficult to be suitable for millions of pixel scenes; (2) The self-expression construction efficiency is low, the existing method is not beneficial to end-to-end optimization due to the fact that the self-expression construction efficiency is low, or the existing method depends on iterative optimization (such as ADMM), the fracture characteristic learning and the self-expression solution, or the self-expression matrix is parameterized into a full-connection layer, and redundant parameters are introduced; (3) The view alignment mechanism is weak, and the mainstream contrast learning method only aligns sample level characteristic representations, so that the subspace structure with consistent bottom is difficult to learn when strong noise or multi-view difference is obvious. Therefore, there is a need for an unsupervised PolSAR image classification method that is scalable, efficient, and can achieve structure level view alignment, so as to break through the practical bottleneck of large-scale remote sensing image processing. Disclosure of Invention In order to solve the above-mentioned key technical bottlenecks of poor expandability, low self-expression construction efficiency, weak view alignment mechanism and the like, the technical problems of high-efficiency, robust and end-to-end optimized unsupervised classification of the existing image classification method are difficult to realize, the invention aims to provide an unsupervised PolSAR classification method based on contrast learning, and the adopted technical scheme is as follows: An embodiment of the invention provides an unsupervised PolSAR classification method based on contrast learning, which comprises the following steps: Acquiring original PolSAR image data, and extracting at least two pixel-level features to form multi-view data; performing superpixel segmentation on the image, and respectively aggregating the characteristics of all pixels in each superpixel under each view to generate a multi-view element sample corresponding to the superpixels one by one; Inputting the multi-view element samples into a corresponding depth feature extraction network to obtain potential feature representation of each view, and generating a self-representation coefficient matrix of each view through a forward self-representation module, wherein each column corresponds to a self-representation coefficient vector of one element sample, and represents the weight of the self-representation coefficient vector linearly reconstructed by all element samples; Based on a relationship consistency contrast learning mechanism, constructing self-expression coefficient vectors of the same element sample under different views into positive sample pairs, applying cross-view relationship consistency constraint through contrast loss, and optimizing a self-expression coefficient matrix of each view; generating a unified affinity matrix by fusing the optimized self-expression coefficient matrix of each view, and performing spectral clustering to obtain a clustering label of the meta-sample; And mapping the clustering label back to the pixel level according to the mapping relation between the meta-sample and the pixel established by the super-pixel segmentation, and obtaining an image classification result. Further, the multi-view data includes a polarization feature view and a texture feature view. Further, the aggregation employs a truncated mean aggregation function. Further, obtaining the self-expression coefficient matrix includes: carrying out L2 normalization on p