An image classification using convolutional sparse representation and cone-restricted subspace method
Yosuke Higuchi, Tomoya Hirakawa, Yoshimitsu Kuroki · 2021
This paper presents a classification method based on Convolutional Sparse Representation (CSR) and Cone Restricted Subspace Method (CRSM). CSR extracts image features as convolutional filters and shows the feature maps as coefficients of the filters. This scheme is similar to a convolutional layer of CNNs (Convolutional Neural Networks). To increase the robustness against image shift, this work uses power spectrum as a non-linear operator like activate functions on CNNs. CRSM approximates each set of class as a cone because of non-negativeness of the power spectrum; furthermore, Nonnegative Matrix Factorization (NMF) generates the basis vectors of CRSM. Experiments on handwritten image classification shows that our method using a 3-layer network shows higher recognition ratio than the same structure of CNN when the number of learning images is less than 1,000.