Commonality Autoencoder: Learning Common Features for Change Detection From Heterogeneous Images.

Area of Science:

• Remote Sensing
• Computer Vision
• Artificial Intelligence

Background:

• Change detection using heterogeneous images (e.g., optical, synthetic aperture radar) is difficult due to significant appearance variations.
• Existing methods often struggle with the inherent differences in data modalities, limiting accuracy.

Purpose of the Study:

• To propose an unsupervised change detection method for heterogeneous images.
• To develop a framework that effectively extracts consistent features and explores commonalities between different image types.
• To improve the accuracy and robustness of change detection in complex scenarios.

Main Methods:

• Utilizing a convolutional autoencoder (CAE) for feature extraction to reduce redundancy and achieve consistent representations.
• Employing a commonality autoencoder to discover shared features between heterogeneous images by transforming representations.
• Learning network parameters based on the relevance of unchanged regions, eliminating the need for labeled data.

Main Results:

• The proposed method demonstrated promising performance across five real-world datasets.
• The commonality autoencoder successfully identified shared features, distinguishing between changed and unchanged regions.
• The generated difference map, after segmentation, yielded accurate change detection results.

Conclusions:

• The unsupervised framework effectively addresses the challenges of heterogeneous image change detection.
• The combination of CAE and commonality autoencoder offers a robust approach for feature extraction and commonality exploration.
• The method shows significant potential for practical applications in remote sensing and other fields requiring change analysis.