ROFANet: Residual Offset-Driven Feature Alignment Network for Unaligned Remote Sensing Image Change Detection

At present, most remote sensing change detection methods are applicable to bitemporal image alignment scenarios, that is, assuming that the pixel pairs of bitemporal images are spatially registered. Detection accuracy is highly sensitive to the alignment accuracy of the image pairs. In practical applications, obtaining well-registered image pairs is often challenging. Currently, the approach of aligning first and then detecting is both inefficient and expensive. Explicitly integrating image alignment and change detection into a framework is an effective solution. However, offset information is difficult to be reflected in the high-level features of the image, it is hard to predict accurate image offsets, and it is also difficult to correct the spatial relationship of land covers in the image using the offset. To overcome the above problems, we propose a residual offset-driven feature alignment network (ROFANet). ROFANet combines two innovative methods: residual offset prediction (ROP) and dual-branch feature correction (DFC). ROP utilizes multilevel features to achieve offset prediction from coarse to fine granularity, effectively enhancing the model's predictive ability for image offsets. DFC has established two branches: image correction and feature correction, which respectively correct distorted images and distorted features. By optimizing the spatial relationship representation of land covers, the model's change detection ability under unaligned image conditions has been enhanced. Extensive experiments conducted on three publicly available change detection datasets demonstrate that the proposed ROFANet achieves outstanding detection performance in unaligned image scenarios.