Deep dilated and densely connected parallel convolutional groups for compression artifacts reduction

Abstract

Abstract A widely used JPEG lossy compression standard introduces different artifacts, particularly, blocking, ringing and blurring effects in the compressed image. Existing algorithms focus on deblocking and discard the ringing, blurring artifacts that produce a visually unpleasant images. These artifacts must be eliminated to restore visually pleasing output while decompressing the image. In this paper, we propose a novel 10-layer network with parallel convolutional groups based on deep dilated convolutions and dense connections (D3C Net) to reduce blocking, ringing and blurring artifacts arise due to JPEG compression. The dilated convolution helps to expand receptive field exponentially without increasing the complexity of the network, thereby relieving the network from the training burden. Moreover parallel convolutional layers are integrated into network for efficient feature representation and model parallelism. In addition, the skip connections allow the information to be back-propagated efficiently to bottom layers and tackles the problem of vanishing gradients. Furthermore, the gradient explosion problem is tackled by an end-to-end mapping between the JPEG decompressed image and the corresponding residual image. Different from the existing algorithms for compression artifacts removal which usually train for a specific compression factor, our D3C Net is able to handle compression artifacts with unknown compression factors (i.e., blind JPEG compression artifacts removal). The subjective and objective performance analysis with metrics PSNR, SSIM and PSNR-B demonstrates that the proposed network produces significant performance improvement when compared to state-of-the-art methods.