Image Pattern Extraction and Compression using Pixel Neighborhood and Weighted PCA Algorithm

Abstract

Background/ Objective: An image can be compressed by compressing the patterns in itself. The patterns preserve the maximum entropy in an image. If patterns are extracted using the pattern extraction algorithm and compressed using their features, then rest of the image may be compressed to a high degree of decompression. The patterns are therefore principal components of the image and principal component analysis can be applied over that in order to achieve an efficient compression algorithm. The principal component analysis algorithm works fine when compressing the input image and outputs a reasonably good compression ratio. However, the compression ratio becomes dependent upon the number of eigen values/vectors chosen to get compressed image. Method: In the presented work, a selection criterion for selecting the Eigen values/vectors is suggested. The criteria is based on threshold selection that is computed by using different techniques and then taking the arithmetic mean of all thresholds from all techniques. Also, a weighted Eigen values are used for threshold computation and compared with the statistical threshold. The Eigen values/vector less than the threshold are taken into consideration for image compression and a well optimized compression ratio is obtained. Finding: The computation time may be improved by decomposing the input image using haar wavelet transform. The different frequency sub bands consist of low frequency (LL), high to low (HL), low to high (LH) and high (HH) frequency sub bands. As the LL sub band contains maximum information out of the four bands, the principal component algorithm is applied in this band only for compression.