Enhancing Low-Resolution Face Recognition with Self Attention Network and Resolution Augmentation Approach

Unconstrained low-resolution face recognition is a very hard and demanding task in the field of computer vision. Due to the scarcity of abundant visual features in low-resolution (LR) images, it is very challenging to extract discriminant key features to improve the identification performance of deep learning networks. The size of the LR images also poses a constraint on the recognition rate, and if it is tiny in size, then it hardens the recognition task more. For boosting the identification rate on unconstrained low resolution faces, a Residual Branch Self Attention Network (RBSaN) is designed. The RBSaN encompasses a self attention module where pixel-wise convolution is applied to channels for extracting the discriminant features. Resolution-augmented images are generated at multiple scales from the high-resolution (HR) images to capture the resolution aspects. RBSaN has achieved identification accuracy on the TinyFace dataset and Surveillance Cameras Face (SCface) database comparable to state-of-the-art (SOTA) models. Heat-map visualization witnesses the effectiveness of RBSaN to locate the key discriminant feature areas in low-resolution images for separating the identities in embedding space. Extensive experimental analysis has achieved qualitative results comparable to SOTA models.