Difference Perception Fusion Network for Remote Sensing Image Change Detection

Abstract

Deep-learning-based methods have achieved promising results in the field of remote sensing image change detection. However, they have deficiencies in generating and utilizing difference features, which leads to inaccurate detection of changed regions with fine structures. To address the aforementioned problems, this article proposes a difference perception fusion network. First, multilevel bitemporal features are extracted through the Siamese backbone network. To generate discriminative difference features utilizing the extracted multilevel bitemporal features, the difference perception module is designed, which consists of an adaptive denoising module (ADM) and an edge enhancement module (EEM). Specifically, ADM relies on frequency filters with learnable parameters to suppress the noise produced during difference generation and EEM uses pooled subtraction to extract edge information for maintaining the fine contours of changed regions. The hybrid attention difference fusion module is constructed to realize the interaction between multilevel difference features, enriching the edge details and internal integrity of changed regions. In addition, the refinement representation module is developed to realize the sophisticated representation of difference features and enhance the detection effect of changed regions. The effectiveness of the proposed DPFNet is verified by extensive experiments on three RSCD datasets (WHU-CD, CDD-CD, and SYSU-CD), and DPFNet has fewer false detections and missed detections compared to other state-of-the-art methods.