基于Kohonen神经网络的分形图像编码

本文提出利用Kohonen自组织神经网络把母块分类与特征抽取结合起来有助于改善分形编码的时间。因为特征抽取减少了问题的维数并且使网络能够在一幅和实验图像分离的图像上得到训练。自组织网络为分类引入了一个领域拓扑结构，并且不需要事先指定一组适当的图像类。网络按照在训练期间观测的图像特征的分布来组织自己。结果表明，该分类方法可以将编码时间减少两个数量级并保持可观的精度和压缩性能。

AN APPROACH TO FRACTAL IMAGE CODING BASED ON Kohonen NEURAL NETWORK

A scheme for improving encoding times in fractal image compression is presented in this paper. The approach combines feature extraction with domain classification by using a Kohonen self-organizing neural network. Feature extraction reduces the dimensionality of the problem and enables the neural network to be trained on an image which is separated from the test image. The self-organizing network introduces a neighborhood topology for classification, and also eliminates the need to specify prior a set of appropriate image classes. The network organizes itself according to the distribution of the image features observed during training. The results show that this classification approach can reduce encoding times by two orders of magnitude, while comparable accuracy and compression performance are maintained.