The paper introduces PCANet, a simple deep learning network for image classification, consisting of basic data processing components: cascaded principal component analysis (PCA), binary hashing, and block-wise histograms. PCA is used to learn multistage filter banks, followed by binary hashing and block histograms for indexing and pooling. The architecture is named PCA Network (PCANet) and is designed and learned efficiently. Two variations, RandNet and LDA-net, are introduced, which share the same topology but differ in their cascaded filters. Extensive experiments on various benchmark datasets, including face recognition, handwritten digit recognition, texture classification, and object recognition, show that PCANet performs on par with state-of-the-art features, often outperforming them. PCANet achieves competitive or superior results in several tasks, such as face recognition in Extended Yale B, AR, and FERET datasets, and handwritten digit recognition in MNIST variations. The simplicity and effectiveness of PCANet make it a promising baseline for future research in deep learning and visual recognition.The paper introduces PCANet, a simple deep learning network for image classification, consisting of basic data processing components: cascaded principal component analysis (PCA), binary hashing, and block-wise histograms. PCA is used to learn multistage filter banks, followed by binary hashing and block histograms for indexing and pooling. The architecture is named PCA Network (PCANet) and is designed and learned efficiently. Two variations, RandNet and LDA-net, are introduced, which share the same topology but differ in their cascaded filters. Extensive experiments on various benchmark datasets, including face recognition, handwritten digit recognition, texture classification, and object recognition, show that PCANet performs on par with state-of-the-art features, often outperforming them. PCANet achieves competitive or superior results in several tasks, such as face recognition in Extended Yale B, AR, and FERET datasets, and handwritten digit recognition in MNIST variations. The simplicity and effectiveness of PCANet make it a promising baseline for future research in deep learning and visual recognition.