Color Image Compression Through Rough PCA and Feed Forward Neural Networks

Abstract

In this paper color image compression using a hybrid neural-network model consisted of rough principal component analysis (RPCA) for separating an image into rough principal components and a feed forward network (FFN) to restore rough principal channels into the original information was proposed. In this system, a training vector is only consisted of two pixels with maximum and minimum gray values in a subdivided block for the input of RPCA net in the encoder. The proposed RPCA generates several channels with black and white image as the original color image but each sample represented by a scale value instead of a three-coordinate vector samples. Then these rough principal components are directly fed into FFN to train the synaptic weights between input and output neurons. Finally, the rough principal components and the synaptic weights are transmitted to the decoder. The experimental results show that the proposed RPCA and feed forward networks can obtain the promising reconstruction performance.