Robustness of PCANet with Block Optimization in Noisy Image Recognition

In noisy image recognition tasks, traditional global feature algorithms (e.g., Principal Component Analysis (PCA), original Principal Component Analysis Network (PCANet)) face challenges of sensitivity to local noise and inefficient training, leading to feature deviation and recognition errors in highrobustness scenarios. This paper presents a Block-Optimized PCANet (BPCANet) by introducing structured non-overlapping blocking strategy to suppress local noise contamination on global features, and a hierarchical feature concatenation fusion mechanism to preserve sub-block spatial location and detailed features, avoiding noise-dominated feature masking. Experiments on Yale and AR noisy face datasets show that BPCANet significantly outperforms PCA, Multilinear Principal Component Analysis (MPCA), and original PCANet in recognition accuracy (e.g., 96.67% accuracy on Yale with 20% noise, <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$+31.82 {\%}$</tex> over PCA) while reducing training time by over 60% compared to original PCANet. The results validate the blocking structure's effectiveness in balancing noise robustness and computational efficiency, offering an optimized solution for image recognition in complex noisy environments.