Adaptive Extraction of Principal Colors Using an Improved Self-Growing Network

Abstract

This paper aims to solve the two major issues existing in current color quantization algorithms. The first one is to require users to specify the number of representative colors in advance; the other is that it is difficult in choosing the colors to describe accurately the essential details represented by small groups of pixels isolated in the color space. Based on the growing mechanism of the Growing When Required neural network, a novel algorithm is proposed to adaptively extract the prominent colors of an image. A number of criteria are introduced that have an effect on controlling of the number and topology of neurons in the output layer. A global permutation method to rearrange the input sample order is presented based on Linear Pixels Shuffling in order to improve the performance of the network. The experiments show that the proposed method can automatically estimate the number of colors to efficiently represent an original image, meanwhile capable of retaining important isolated colors even when the number of the representative colors is low. It is also shown that the algorithm outperforms the popular ones in terms of color distortion.