HQFS: High-Quality Feature Selection for Accurate Change Detection.

Yan Xing,Qi'ao Xu,Qingyi Zhao,Rui Huang,Yuxiang Zhang

DOI: https://doi.org/10.1007/978-3-031-46305-1_3

2023-01-01

Abstract:High-quality features play a crucial role in achieving accurate change detection (CD). However, current CD methods often rely on simple combinations of features or the computations of feature differences between bi-temporal images, which cannot fully explore the representative capacity of features. In this paper, we design a high-quality feature selection module , called HQFS, to generate high-quality change-related features for accurate CD. Firstly, we analyze the pros and cons of the features utilized in recent deep learning-based CD methods. Based on this analysis, we construct a feature pool containing various change-related features. The proposed HQFS module is then employed to learn the importance of these features and assign higher weights to the high-quality features to amplify their contributions in computing the change maps. To further enhance the performance, we propose two additional components: the fully attention enhanced feature extraction network and the attention pyramid fusion & change mask prediction module . The former focuses on extracting features from images, while the latter leverages attention mechanisms and pyramid fusion techniques to generate accurate change predictions. Comprehensive experiments on three benchmark datasets demonstrate the superiority of our method over seven state-of-the-art change detectors.