Correlation and Attention Mechanism based Network for Change Detection

Change detection has experienced substantial progress in computer vision. Recently, siamese structure based deep neural networks have been successfully applied for change detection problem. Yet, there still exist two obstacles for addressing this problem on bi-temporal images. 1) The compared images are collected on a moving platform like car or other robotics at different times, which leads to the generation of roughly aligned pairs. 2) There exist various type noises like disturbing noise and changing light, which easily lead to rough boundaries and large holes in real applications. Therefore, a change detection framework named dual correlation attention mechanism network (DCAMNet) is designed for this detection task in this paper. Firstly, for the first issue, correlation units are introduced into an end-to-end framework to align the paired images through feature representations. Secondly, for the season changes, noise disturbance, blurred edges problems, spatial-temporal units are designed to provide high-quality shapes and change regions. Finally, extensive experimental results on the "CDnet 2014 dataset" and "LEVIR-CD dataset" show promising performance of our introduced method.