Object Relations Focused Siamese Network for Remote Sensing Image Change Detection

AbstractIn recent years, deep learning methods, especially convolutional neural networks (CNNs), have shown powerful discriminative ability in the field of remote sensing change detection. However, the multi-temporal remote sensing images often have a long time interval, which leads to the geographical objects with the same semantic may have different spectral representations in different spatial and temporal locations, thus affecting the performance of network detection. Exploring the latent connections between different objects helps promote the discrimination of feature maps, so as to suppress the generation of pseudo changes because the geographical objects are not isolated. In this paper, we propose a novel change detection network focused on the object relations (ORFNet), which can capture relevant context by exploring the relations between objects in dual spatio-temporal scenes, so as to enhance the discrimination of original image features. Experiments on CDD data sets show that our method only increases 0.4M parameters compared with the baseline, and improves F1 by 3.6%.KeywordsDeep learningConvolutional neural networksObject relationsFeature enhancement