A dual-residual network for JPEG compression artifacts reduction
- PDF / 1,111,515 Bytes
- 7 Pages / 595.276 x 790.866 pts Page_size
- 87 Downloads / 202 Views
ORIGINAL PAPER
A dual-residual network for JPEG compression artifacts reduction Jianfei Li1 · Dongsheng Li1 · Chunxiao Chen1 · Qiang Yan1 · Xiong Lu2 Received: 8 October 2019 / Revised: 8 August 2020 / Accepted: 17 August 2020 © Springer-Verlag London Ltd., part of Springer Nature 2020
Abstract Current image compression techniques such as JPEG and WebP are widely applied in the age of information. However, lossy compression like JPEG in its nature will introduce visually annoying artifacts while saving Internet bandwidth and storage space. The artifacts such as blocking and ringing are especially sharp at low bitrates. In this paper, we propose a novel dual-residual network to reduce compression artifacts caused by lossy compression codecs. This network directly learns an end-to-end mapping between the distorted image processed by JPEG or other compression methods and the original image, which takes decompressed images with blocking artifacts as input and produces clearer images with less artifacts. Experiments results on the test dataset demonstrate the efficiency of the proposed model, especially at very low bitrates. Keywords Dual-residual network · Image compression · JPEG · Artifacts reduction
1 Introduction In recent years, image and video compression is playing a more and more important role in reducing the amount of multimedia data both on the Internet and offline. It is necessary and essential to reduce irrelevance and redundancy within an image or video so as to transmit and store data at less costs [1]. For example, JPEG, as a traditional image compression standard [2], is still applied in many fields nowadays as NASA using JPEG to process Mastcam multispectral images [3]. JPEG is conducted based on image blocks with the pixel block size of 8 × 8. Discrete cosine transformation (DCT) and quantization for each independent block are applied during the encoding process in order to distribute fewer bits for the whole image, and compression ratio increases as quantization steps for DCT coefficients become larger. The quantization step is determined by quality factor (QF) [4], and small QF indicates large step as well as low quality of decoded images. The results of peak signal-to-noise ratio (PSNR) for JPEG decompressed image and that processed by our model are shown in Fig. 1. We can find that ringing
B 1
2
and blurring artifacts in the JPEG-decoded image with QF 5 tend to be more apparent, which greatly restrict the use of compressed low-level images. The traditional artifacts-depressing algorithms aim at removing blocking and ringing regions by means of deblocking and denoising filters or generating more details to restore images. The former would cause serious blurring while the later can bring in undesired noises and unnatural edges. Recently, deep learning has achieved great success for both high-level and low-level vision tasks. Experiments have proved the impressive results of neural networks in dealing with problems in the field of computer vision, such as image recognition [5, 6], image segm
Data Loading...
