The Concatenated Dynamic Convolutional and Sparse Coding on Image Artifacts Reduction

Linna Yang, Ronny Velastegui

Abstract: In order to enhance compressed JPEG image, a deep convolutional sparse coding network is proposed in this article. The network integrates state-of-the-art dynamic convolution to extract multi-scale image features, and uses convolutional sparse coding to separate image artifacts to generate coded feature for the final image reconstruction. Since this architecture consolidates model-based convolutional sparse coding with deep neural network, that allow this method has more interpretability. Also, compared with the existing network, which uses a dilated convolution as a feature extraction approach, this proposed concatenated dynamic method has improved de-blocking result in both numerical experiments and visual effect. Besides, in the higher compressed quality task, the proposed model has more pronounced improvement in reconstructed image quality evaluations.

External IDs:dblp:conf/iccsa/YangV21