Performance analysis of JPEG XR with deep learning-based image super-resolution

Abstract

The demand for efficient high-level image and video codec compression has widely increased. Conventional image compression methods such as JPEG XR use a high quantization parameter (QP) to produce a highly compressed file for any given image. However, higher QP has unpleasing artifacts that lead to perceptual quality degradation. A feasible solution to tackle this limitation is to reduce the high-resolution image size by downsampling it before encoding it with JPEG XR. Then, the super-resolution algorithm is applied to the resultant low-resolution image to reconstruct the high-resolution result. In this research, we downsample the input image before JPEG XR. Then, we investigate the performance of integrating a newly retrained deep learning-based FSRCNN super-resolution (SR) with JPEG XR in terms of quality and compressed file size. According to the experimental results, the experimental results show that the proposed method outperforms JPEG XR compression by shrinking the size of the encoded file by an average of 557 kB for scale two and 756 kB for scale four. The fusion of the newly trained model with JPEG XR compression can achieve higher performance than JPEG XR compression in compressing the file, around 66% for scale two and 89% for scale four. The proposed method also produces a small, compressed file size for high compression and achieves better visual quality than JPEG XR compression, JPEG XR with the bicubic method, and JPEG XR with FSRCNN.