Learning Block Group Spase Representation Combined with Convolutional Neural Networks for RGB-D Object Recognition

Abstract

RGB-D (Red, Green and Blue-Depth) cameras are novel sensing systems that can improve image recognition by providing high quality color and depth information in computer vision. In this paper we propose a model to study feature representation of combined Convolutional Neural Networks (CNN) and Block Group Sparse Coding (BGSC). Firstly, CNN is used to extract low-level features from raw RGB- D images directly by applying unsupervised algorithm. Then, BGSC is used to obtain higher feature representation for classification by incorporating both the group structure for low-level features and the block structure for the dictionary in subsequent learning processes. Experimental results show that the CNN-BGSC approach has higher accuracy on a household RGB-D object dataset by linear predictive classifier than using Convolutional and Recursive Neural Networks (CNN-RNN), Group Sparse Coding (GSC), and Sparse Representation base Classification (SRC).