End-to-end image compression method based on perception metric

In recent years, image compression methods based on deep learning have received extensive attention and research. Most methods focus on minimizing the mean squared error (MSE) to obtain reconstructed images with higher peak signal-to-noise ratio (PSNR). However, the ability of pixel-wise distortion to capture the perceptual differences between images is fairly limited, which may suffer from undesirable visual perception quality of the reconstructed images. To address this problem, we propose a novel rate-distortion loss based on perception metric in learned image compression. In this work, we introduce the perception metric into the rate-distortion loss, which can enhance the capacity of compression model to capture perceptual differences and semantic information in images. By performing that, the rate-distortion performance of our proposed model on multi-scale structural similarity (MS-SSIM) and the classification accuracy of reconstructed images have been improved. Comprehensive experimental results demonstrate that the proposed method has comparable performance in terms of PSNR, and the performance on MS-SSIM outperforms traditional image codecs, such as JPEG and BPG, as well as other previous end-to-end compression methods. More significantly, the visual quality of the reconstructed images is dramatically improved.