Adjacent paired LBP and enhanced HOG for improved facial expression recognition

Facial expression recognition remains a challenging problem in computer vision due to subtle variations in facial features and complex background conditions. In this paper, we propose a novel method, named Adjacent Paired Local Binary Pattern (AP-LBP), to enhance the discriminative power of local texture features. Unlike conventional LBP, which only compares each neighboring pixel with the central pixel, AP-LBP encodes the structural relationships among neighboring pixels themselves. Specifically, within a 3×3 patch, AP-LBP performs two rounds of comparisons: first between adjacent pixel pairs in a clockwise sequence, and then between pixel pairs with a stride of one. The two binary codes are fused via averaging to form the final representation. This design allows AP-LBP to capture richer contextual information without enlarging the local receptive field or increasing computational complexity. To further improve feature representation, we also present an enhanced Histogram of Oriented Gradients (HOG) descriptor by incorporating diagonal gradients in addition to standard horizontal and vertical components, yielding better orientation sensitivity. Experimental results on standard facial expression datasets demonstrate that the proposed method achieves superior recognition performance compared to LBP and HOG descriptors.