Semi-NMF network for image classification

Semi non-negative matrix factorization (Semi-NMF) is an algorithm that gets a low-dimensional representation of a database. In this paper, we propose an efficient convolutional neural network (CNN) based on Semi-NMF to address the image classification problem. Unlike the traditional learning approach of CNN, convolutional filters are achieved by the Semi-NMF on input images. In the output layer, binary hashing and blockwise histograms are used to obtain the feature maps. To get better classification results, we introduce the weakly supervised Semi-NMF network which incorporates known attribute information for computing the convolutional filters in first layer. Experiments on MNIST dataset show our methods perform better than the state-of-the-art methods.