Improving Quantization Matrices for Image Coding by Machine Learning

Abstract

We investigate the generation of quantization matrices for image coding in the scenario to balance compression ratio and quality. We make use of machine learning to train and determine those quantization matrices that can achieve the best compression ratio while reaching the quality settings. By introducing the trainable parameters and considering the impact of the quantization module on task performance and compression ratio, the DCT and quantization modules are jointly optimized to minimize the total coding cost. We evaluate the well-trained quantization matrices under various quality settings of JPEG. The results indicate that the proposed scheme can be combined with quality settings to consistently achieve better compression performance.