Local and global feature attention fusion network for face recognition

Yu Wang, Wei Wei

Abstract: Highlights•We categorize the factors leading to low-quality facial images into three categories , and the quality of local and global features extracted from different categories of face images are biased. We introduce the LGF module, which dynamically fuses local and global features based on attention for the first time, optimizing facial representation and mitigating feature quality biases.•We explore, for the first time, the relationship between the norm of local and global features and different categories of low-quality facial images. We propose that the norm of local features is highly correlated with the degree of deformation in facial regions, while the norm of global features is more sensitive to the extent of facial missing. Consequently, we suggest that feature norm can serve as proxy for feature quality.•We optimize the local feature extraction network by acknowledging that local features at different scales exhibit varying spatial distributions. We propose the MHMS module, which extracts spatial attention for local feature maps at each scale separately, enabling the model to capture more comprehensive local information.