Feature Enhancement and Reweighting for Transformer-Based Change Detection

As Transformer is more widely used in the domain of Computer Vision (CV), modern techniques for Change Detection (CD) have also begun to use Transformer structures, including Bitemporal Image Transformer (BIT). Although BIT shows excellent performance due to its efficient context modeling ability, the simple backbone network and the Cross-Entropy (CE) loss it uses still have room for improvement. In this paper, we propose a Feature Pyramid Network of Change Detection (FPN-CD) and a Change Detection focal (CDF) loss to address the shortcomings of the BIT method. Meanwhile, the outcomes of ablation experiments performed on two CD datasets attest to the method's efficacy.