基于神经网络和线性像素置换的颜色量化

Color quantization algorithm based on neural network and linear pixel shuffling

本文提出了一种新的彩色图像量化算法。它是一种基于自组织神经网络和线性像素置换的后聚类算法。线性像素置换是一种均匀选取图像中的像素的方法。根据线性像素置换确定改进的自组织神经网络的初始权重向量和训练样本集。选取部分样本参加训练加快训练过程。实验结果表明,与其它量化优化算法比较,本文提出的算法在量化图像质量和算法效率方面均有明显提高,而且不依赖于算法的初始条件。

In this paper, a novel color quantization algorithm is presented. It is a post-clustering technique, based on Self-Organizing Kohonen Network and Linear Pixel Shuffling （LPS）. LPS provides a method for uniformly visiting pixels in an image. The initial weighted vectors and training sets are also determined by LPS. Limited samples are taken in order to speed up the training process of the improved neural network. The influence of different values of sampling rate is discussed. The presented algorithm is compared with other well-known approaches in terms of quantization error, executive time as well as human perception. Experiments show that the proposed algorithm results in a significant improvement of image quality and reduction of the rtmning time without depending on the set of initial conditions