Fast and Accurate Color Image Classification Based on Quaternion Tchebichef Moments and Quaternion Convolutional Neural Network

This paper introduces a new architecture named QTMCNN for color image classification based on quaternion discrete Tchebichef moments (QTM) and quaternion convolutional neural network (QCNN) to improve the classification accuracy and to reduce the time of learning process. Color image is represented as a single quaternion matrix where each color pixel is represented as a pure quaternion. From this representation, quaternion Tchebichef moments are used to generate a matrix of low-dimensional significant features and fed to QCNN as input layer instead of color image. The proposed architecture reduces tremendously the number of parameters and consequently decreases the computational complexity while improving the classification rates. Experiments are conducted on Coil-100 and ETH-80 datasets to demonstrate the performance of the proposed architecture. The obtained results outperform other approaches in terms of classification accuracy and GPU elapsed time.