Efficient feature coding based on auto-encoder network for image classification

Abstract

Local descriptor coding is one crucial step in traditional Bag of Words (BoW) framework for image categorization. However, the slow coding speed of previous methods is one limitation for applications in large scale problems. Recently, neural network based models have been widely applied in various classification tasks. Using neural network models for descriptor coding is straightforward and efficient due to their fast forward propagation. In this paper, we propose to use the Auto-Encoder (AE) network as a local descriptor coding block, and further embed AE network in the BoW framework for the purpose of image classification. To make the hidden activities of AE network to be both selective and sparse, we add an efficient and effective regularization term into the learning process of AE network, which can promote sparsity of the hidden layer for each input descriptor as well as the selectivity for each hidden node. By incorporating the AE network coding with the BoW framework, we can achieve better results and faster speeds than other state-of-theart feature coding methods on Caltech101, Scene15 and UIUC 8-Sports databases.