A Prediction Scheme for Image Vector Quantization Based on Mining Association Rules

Abstract

Vector Quantization (VQ) is an efficient method for image compression. Many conventional VQ algorithms for lower bit rates, such as SMVQ, consider only adjacent neighbors in determining a codeword. This leads to awful distortion. In this paper, we propose an efficient association rules mining method inspired by an approach widely adopted in data mining, for predicting image blocks based on the spatial correlation. The proposed method is divided into two parts. First, it generates dominant vertical, horizontal, and diagonal association rules of training images. Then it searches for a suitable replacement according to the matched rules. The rule-based method for prediction is more efficient than conventional VQ since finding the matched rules is easier than calculating the distances between codewords. The experimental results show that our method is excellent in the performance in terms of both image quality and compression rate.