CN-121982560-A - Micro-filter combined automatic searching heterogeneous remote sensing image change detection method and system

CN121982560ACN 121982560 ACN121982560 ACN 121982560ACN-121982560-A

Abstract

The invention discloses a method and a system for detecting the change of a micro-filter combined automatic searching heterologous remote sensing image, wherein the method comprises the following steps of S1, carrying out feature fusion processing on two obtained heterologous remote sensing images to obtain fusion features; S2, dynamically synthesizing convolution kernels in the discriminators by adopting a micro-filter combined search strategy according to the fusion characteristics, optimizing parameters of the generator and the discriminators by countermeasure training, and S3, supervising a variation map output by the generator based on geometric structure collaborative supervision loss to obtain a final variation detection result. The method can effectively extract the common texture and structural characteristics in the optical and SAR images, overcomes the problems of large radiation difference and high noise of the heterogeneous images, and obviously reduces the omission ratio and the false detection ratio. Meanwhile, the invention cooperates with pixel precision, edge definition and global structure authenticity, so that the detected change area has clear boundary and complete structure, and common salt and pepper noise and boundary blurring phenomena are effectively eliminated.

Inventors

ZHANG YAN
LI HUI
CHEN JIE
GUO XIANFEI
Wan Huiyao
Pazilaiti Nurmaiti
XU YUNHENG
HUANG ZHIXIANG
YANG ZHIGAO

Assignees

安徽大学
中科卫星(安徽)数据科技有限公司

Dates

Publication Date: 20260505
Application Date: 20251230

Claims (8)

1. The method for detecting the change of the micro-filter combined automatic searching heterologous remote sensing image is characterized by comprising the following steps of: s1, performing feature fusion processing on two acquired heterologous remote sensing images to obtain fusion features; s2, dynamically synthesizing convolution kernels in the discriminators by adopting a micro-filter combined search strategy according to the fusion characteristics, and optimizing parameters of the generator and the discriminators through countermeasure training; And S3, supervising the variation map output by the generator based on the geometric structure collaborative supervision loss to obtain a final variation detection result.
2. The method for detecting the change of the automatically searched heterogeneous remote sensing image by using the micro-filter combination according to claim 1, wherein the step S1 comprises: Gabor filtering processing is carried out on the two heterologous remote sensing images respectively, and a multidirectional texture feature map of each image is obtained; Calculating a local normalized cross-correlation map between the two images according to the multi-directional texture feature map; and splicing the original convolution feature, the multi-directional texture feature map and the local normalized cross-correlation map to obtain a fusion feature.
3. The method for automatically searching for a change in a heterologous remote sensing image by a combination of microfilters according to claim 2, wherein the method for calculating a locally normalized cross-correlation profile between two images comprises: respectively calculating the average value of the features in the feature windows at the corresponding positions in the two images; and calculating cosine similarity between the two centralized feature vectors based on the mean value of the features in the window and the feature data of each point in the window to obtain a local normalized cross-correlation value.
4. The method for detecting the change of the automatically searched heterogeneous remote sensing image by using the micro-filter combination according to claim 1, wherein in S2, the searching strategy by using the micro-filter combination comprises: Defining a set of learnable architecture parameters for each searchable convolutional layer in the arbiter; Converting the architecture parameters into normalized combining weights by a Softmax function; and linearly combining the combination weight with a preset base filter to generate a dynamic convolution kernel.
5. The method for automatically searching for a change in a heterologous remote sensing image based on a combination of microfilters according to claim 4, wherein S2 further comprises: constructing a micro-filter combination search as a double-layer optimization problem, wherein an upper-layer optimization target is used for searching an optimal architecture parameter, and a lower-layer optimization target is used for searching an optimal network weight under a given architecture parameter; By alternately updating the network weight and the architecture parameters, the adaptive optimization of the structure of the discriminator is realized.
6. The method for automatically searching for a change in a heterologous remote sensing image based on a combination of microfilters according to claim 1, wherein the geometrical collaborative supervision loss in S3 comprises a weighted sum of pixel-level region loss, edge consistency loss and global structure countermeasure loss.
7. The method for automatically searching for a change in a heterologous remote sensing image based on a combination of microfilters according to claim 6, wherein the pixel-level regional loss comprises a combination of binary cross entropy loss and cross-ratio loss; the edge consistency loss is obtained by calculating the L1 distance between the generated change graph and the truly marked gradient graph; The global structure countermeasures against loss based on the discrimination results of the discriminator on the real change map and the generated change map, and aims at maximizing the probability that the generated change map is discriminated as the real change map, and a generator loss function is constructed.
8. A micro-filter combined automatic searching heterogeneous remote sensing image change detection system for realizing the method of any one of claims 1-7, comprising a fusion module, an countermeasure training module and a detection module; The fusion module is used for carrying out feature fusion processing on the two acquired heterogeneous remote sensing images to obtain fusion features; The countermeasure training module is used for dynamically synthesizing convolution kernels in the discriminators by adopting a micro-filter combination search strategy according to the fusion characteristics and optimizing parameters of the generator and the discriminators through countermeasure training; the detection module is used for supervising the variation map output by the generator based on the geometric structure collaborative supervision loss to obtain a final variation detection result.

Description

Micro-filter combined automatic searching heterogeneous remote sensing image change detection method and system Technical Field The invention relates to the technical field of remote sensing image processing and deep learning, in particular to a method and a system for detecting the change of a remote sensing image capable of automatically searching for a different source by a micro-filter combination. Background The heterologous change detection (Hete-CD) of the remote sensing image aims at identifying the change of the ground object between images acquired by different sensors (such as optics and SAR), and has important value in the fields of disaster assessment, urban monitoring and the like. However, het-CDs face a great challenge: the large modal differences, the imaging mechanisms of different sensors (e.g., optics and SAR), result in significant differences in the geometry, radiation characteristics, and noise distribution of the image (i.e., "semantic gaps"), making it extremely difficult to directly compare pixels. Network architecture design is limited by existing deep learning (especially GAN) based methods, typically employing a fixed manually designed arbiter architecture. The fixed architecture often has experience deviation, is difficult to adapt to complicated and changeable heterogeneous data distribution, and leads to the fact that the model is easy to sink into local optimum. Boundary and structure blurring-conventional loss functions (such as cross entropy) typically only focus on pixel-level classification accuracy, neglecting the structural integrity and spatial continuity of the varying regions, resulting in blurring of the generated detection map boundary, and susceptibility to salt and pepper noise. In the prior art, although methods based on image conversion or feature projection exist, when processing heterogeneous data in a complex scene, it is often difficult to simultaneously combine noise suppression, detail preservation and feature alignment, so that detection accuracy and F1 score are limited. Therefore, a new approach is needed that can adaptively extract modality invariant features and automatically optimize the network architecture to provide a more strongly supervised signal. Disclosure of Invention The invention provides a self-adaptive anti-learning heterogeneous change detection method named as DFCD-Net. The method aims to solve the problems of difficult characteristic alignment, fuzzy boundary and weak generalization capability caused by large modal difference and fixed network architecture in the prior art. The invention enables the arbiter to dynamically synthesize the optimal convolution kernel by constructing a micro-filter combined search strategy (DFCS), provides self-adaptive countermeasures and supervision for the generator, combines a feature fusion module based on Gabor and local normalized cross-correlation (G-LNCC) to effectively extract the mode invariant features, and utilizes the geometric structure collaborative supervision loss (GSCS) to carry out refined constraint on the change map from three dimensions of pixels, boundaries and structures so as to realize high-precision heterogeneous image change detection. In order to achieve the above purpose, the invention provides a method for detecting the change of a micro-filter combined automatic searching heterologous remote sensing image, which comprises the following steps: s1, performing feature fusion processing on two acquired heterologous remote sensing images to obtain fusion features; s2, dynamically synthesizing convolution kernels in the discriminators by adopting a micro-filter combined search strategy according to the fusion characteristics, and optimizing parameters of the generator and the discriminators through countermeasure training; And S3, supervising the variation map output by the generator based on the geometric structure collaborative supervision loss to obtain a final variation detection result. Preferably, the S1 includes: Gabor filtering processing is carried out on the two heterologous remote sensing images respectively, and a multidirectional texture feature map of each image is obtained; Calculating a local normalized cross-correlation map between the two images according to the multi-directional texture feature map; and splicing the original convolution feature, the multi-directional texture feature map and the local normalized cross-correlation map to obtain a fusion feature. Preferably, the method for calculating the local normalized cross-correlation map between two images comprises: respectively calculating the average value of the features in the feature windows at the corresponding positions in the two images; and calculating cosine similarity between the two centralized feature vectors based on the mean value of the features in the window and the feature data of each point in the window to obtain a local normalized cross-correlation value. Preferably, in S2, the microf