Jpeg Image Super-Resolution Via Deep Residual Network
INTELLIGENT COMPUTING METHODOLOGIES, ICIC 2018, PT III(2018)
摘要
In many practical scenarios, the images to be super-resolved are not only of low resolution (LR) but also JPEG compressed, while most of the existing super-resolution methods assume compression free LR image inputs. As a result, the JPEG compression artifacts (e.g., blocking artifacts) are often exacerbated in the super-resolved images, leading to unpleasant visual results. In this paper, we address this problem via learning a deep residual convolutional neural network (CNN) that exploits a skips-in-skip connection. More specifically, by increasing the network depth to 31 layers with receptive field of 63 by 63, we train a single CNN model which is able to handle JPEG image super-resolution with various combinations of scale and quality factors, as well as the extreme cases, i.e., image super-resolution with multiple scale factors, and JPEG image deblocking with different quality factors. Our extensive experimental results demonstrate that the proposed deep model can not only yield high resolution (HR) images that are visually more pleasant than those state-of-the-art deblocking and super-resolution methods in a cascaded manner, but also deliver very competitive results with the state-of-the-art super-resolution methods and JPEG deblocking methods in terms of quantitative and qualitative measures.
更多查看译文
关键词
JPEG image Super-Resolution, Deep residual network, Skips-in-skip connection
AI 理解论文
溯源树
样例
生成溯源树,研究论文发展脉络