Enhanced Face Recognition With Residual Networks and Dilated Contextual Transformer Attention

Abstract

Face recognition in unconstrained environments remains a challenging problem due to significant variations in facial appearance. In this article, a systematic investigation is conducted into the effectiveness of integrating attention mechanisms within a ResNet34 backbone. ResNet34 is adopted as a compact backbone compared to commonly used deeper ResNet variants, offering a good balance between representational capacity and computational cost. Specifically, three variants are examined: self-attention (SA), contextual transformer (CoT) attention, and the proposed dilated CoT (DCoT) attention. The primary contribution of this study is the demonstration that DCoT, by enlarging the receptive field without substantially increasing the number of parameters, constitutes a more suitable architectural choice for this backbone. This integration enables the network to capture broader contextual information and attend more effectively to discriminative facial features. The proposed models are trained on the CASIA-WebFace dataset and evaluated on standard benchmarks, including labled faces in the wild (LFW), CPLFW, and AgeDB_30. Extensive experiments show that the DCoT-enhanced model achieves consistent improvements over the baseline and other attention variants. Notably, the best-performing model attains an accuracy of 99.60% on the LFW dataset. These findings provide meaningful insights into the interaction between attention mechanisms and convolutional architectures, illustrating the efficacy of the proposed approach for face recognition in unconstrained, real-world conditions, while using a moderately sized backbone that balances capacity and parameter count.