Enhancing the Quality of the BPG Compressed Image

Recent deep learning-based image compression techniques have outperformed the hand-crafted ones like HEVC standard-based Better Portable Graphics (BPG). This codec has a good balance between rate-distortion performance and complexity but is still not as widely adopted as JPEG. This paper investigates the possibility of improving BPG with post-processing to make it more efficient compared to both the old and new compression techniques. For this, we first start by proposing a convolutional neural networks (CNN) model to improve the quality of the BPG images. This model utilizes a UNet architecture, integrating spatial and frequency blocks to enhance feature extraction for image reconstruction. The results demonstrate that this proposed network outperforms many of the state-of-the-art enhancement models. Then we used it to show that the BPG codec, when combined with enhancement, can reduce the efficiency gap between it and the much more complex deep learning-based compression models. Specifically, it improves the PSNR quality by 0.64 dB on the Kodak dataset and 0.78 dB on the Tecnick dataset, making it even better than some learned image compression propositions in terms of rate-distortion performance.