Ensemble leaning-CNN for reducing JPEG artifacts

In this paper, we considered the problem of reducing the artifacts that result when images are compressed by the JPEG encoder. In the last years, many deep learning based techniques, for reducing the JPEG artifacts, appeared. None of these techniques gives always the best quality. In addition, the results depend on the image content and the used quantization. In our work, we proposed to improve the results of existing techniques by an ensemble learning method. Our goal is to obtain the best result for most of the images. For this, we have combined three of the state-of-the-art techniques namely DnCNN, MwCNN, and RidNET. The results of artifacts reduction show a high performance of the proposed network, compared to the individual networks. Thus, it improves the quality PSNR with a gain ranging from 0.01 dB up to 0.34 dB.