Improvement of Image Compression Performance by Deep Neural Networks

This paper analyses the deep image compression system that contains an encoder, quantizer, entropy model and decoder optimized by a joint rate-distortion framework. The current model implements a channel-level variable quantization network to dynamically allocate and withdraw the bitrates from significant and negligible channels. Its main specific is the usage of the variable quantization controller that consists of such components: channel importance module that dynamically learns the importance of channels during the training, and splitting-merging module, which allocates different bitrates of the channels. Quantizer implements the Gaussian mixture model manner. The paper is a continuation of several similar research done before that first provided an idea and architecture of the model. The main goals of the proposed work are to do a deeper analysis of the system, verify and improve the model effectiveness. The experiments validate that the hyper-parameter tuning approach proposed here successfully improves the rate-distortion performance of image compression in terms of various quality metrics such as PSNR, MS-SSIM and BPP.