Reduction of JPEG compression artifacts based on DCT coefficients prediction

Abstract

The image compression-decompression process causes image quality degradations such as blocking and ringing artifacts. A convolutional neural network based on DCT domain is proposed to learn the mapping relationship between JPEG images and original images to reduce compression artifacts in this work. The convolutional neural network can exploit the prior knowledge of JPEG compression in DCT domain. The compression artifacts reduction network which is proposed in this work has three advantages. First, it can expand the receptive field and eliminate the discontinuity between each 8 × 8 block by overlapping extraction of patches. Second, it is based on the essence of distortion and can predict true DCT coefficients more accurately to reduce the loss of JPEG images. Third, it can obviously reduce the compression artifacts in JPEG images. Experiments on compressed images demonstrate that our approach achieves state-of-the-art performance in both the objective parameters and the subjective visual quality.