使用K－L变换进行图象压缩的一种新方法

在图象变换编码领域，Ｋ—Ｌ变换是最小均方误差意义上的最佳变换，但是变换矩阵随图象内容而不同，且计算复杂，速度慢。本文选择了合理的计算自协方差矩阵的特征矢量的方法，以削减计算时间。同时提出将模式识别引入Ｋ—Ｌ变换的一种新的图象压缩的方法，用精选的模式集训练ＢＰ神经网络，使之在计算中将各个子图象正确归类，以选择合适的变换矩阵。这一方法成功地降低了计算复杂性，并且回避了病态矩阵问题。它具有高压缩比和低复杂性的特点。

A New Method for Optimal KLT Image Compression

The opimal linear transformation for image coding with respect to minimizing the mean square error(MSE) is the Karhunen-loeve transformation(KLT). However, KLT matrix specializes to the processed images, and the computation is quite slow and expensive.In this paper, a well-formaed method to perform adaptive calculation of the eigenvectors of the covariance matrix is proposed to reduce computing time, and a new approach of K-L transforms using pattern recognition is proposed. A set of visual patterns was designed as sample set to train a BP network. The algorithm use the trained network to recognize the pattern number of a block image, and use the corresponding metrix to compress the image.The algorithm is successful to solve the computational complexity problem and to avoid the ill-conditioning of the covariance matrix. It offers high compression ratio and low complexity.