Detecting object-level scene changes in images with feature alignment

The change detection task aims to identify changes in data caused by human or natural factors. Although significant progress has been made in deep learning-based methods, existing object-level change detection approaches still face challenges regarding detection accuracy and handling textureless regions, which leads to insufficient precision in detecting these areas during the change detection process. To address these limitations, an object-level change detection network, RCFNet, is proposed. First, based on the Siamese U-Net, RCFNet incorporates a feature refinement module using depthwise separable convolutions within the skip connections to enhance the encoder’s feature representations. In addition, a feature alignment module is introduced, combining attention-based matching with local optical flow estimation. This module consists of two components: self-global cross attention (SGCA) and the semantic flow alignment module (SFAM). The SGCA module strengthens the feature representation of textureless regions and facilitates image registration through cross-image information interaction, whereas the SFAM module explicitly corrects image misalignments to reduce registration errors. The proposed method achieves precise object-level change detection without relying on pixel-level prediction or dense 3D models. Extensive experiments on public datasets such as COCO-Inpainted and VIRAT-STD demonstrate superior performance, particularly in scenarios involving varying viewpoints.