1 ISO/IEC JTC1/SC29/WG1 N4548 Date: March 31, 2008 ISO/IEC JTC 1/SC 29/WG 1 (ITU-T SG16) Coding of Still Pictures JBIG JPEG Joint Bi-level Image JointPhotographic Experts Group Experts Group TITLE: Proposal for objective distortion metrics for AIC standardization SOURCE: J. J. Hwang, S. G. Cho(hwang, [email protected], Kunsan National Univ.), Y. Huh([email protected], KERI) PROJECT: AIC STATUS: Proposal REQUESTED ACTION: For discussion at WG1 Meeting, March, 2008, San Francisco DISTRIBUTION: WG 1 Contact: ISO/IEC JTC 1/SC 29/WG 1 Convener - Dr. Daniel T. Lee Yahoo! Asia, 10 Hysan Avenue, Sunning Plaza, Ste. 2802, Causeway Bay, Hong Kong Yahoo! Inc, 701 First Avenue, Sunnyvale, California 94089, USA Tel: +1 408 349 7051/+852 2882 3898 , Fax: +1 253 830 0372, E-mail: [email protected] 1. Introduction The goal of image quality assessment is to accurately measure how much the reconstruction image is degraded by comparing with its original and utilize the results to design an optimum image codec. Widely used assessment metrics for objective quality measurement are Peak Signal-to-Noise Ratio (PSNR) and Mean Square Error (MSE). These metrics have two drawbacks. First, only one global distortion is calculated so that any specific artifacts may not be measured. Second, they are not considering the human visual properties that should be observed by human viewer. Recently, Structural SIMilarity (SSIM) index is proposed, under the assumption that the Human Visual System (HVS) is highly adapted for extracting structural information from a scene. However, it also does not detect individual artifacts, but provides a global ranking [Wang2004]. Compression and transmission of digital image and video entail a variety of characteristic artifacts and distortions, e.g., blockiness, blurriness, ringing [Yuen1998]. Some of them may be caused by the compression algorithm, while the others as a consequence of transmission errors and various video conversions. Typical artifacts are: • Blockiness or the blocking effect refers to a block pattern in the compressed sequence. It is due to the independent quantization of individual blocks (usually 8 × 8 pixels) in block-based DCT coding schemes. • Blur or the blurriness effect is characterized by the loss of fine detail and the smearing of edges in the video. It is typically caused by high-frequency attenuation at some stage of the recording or encoding process. Wavelet-based encoders also cause blurry artifacts. • Ringing is fundamentally associated with Gibbs’ phenomenon and is thus most evident along high-contrast edges in otherwise smooth area. It is a direct result of quantization leading to high-frequency irregularities in the reconstruction. Ringing occurs with both luminance and chroma components. • False edges are a consequence of the transfer of block-boundary discontinuities (due to the blocking effect) from reference frames into the predicted frame by motion compensation (MC). • Color bleeding is the smearing of the color between areas of strongly differing chrominance. It results from the suppression of high-frequency coefficients of the chroma components. Due to chroma sub-sampling, color bleeding extends over an entire block. • Aliasing can be noticed when the content of the scene is above the Nyquist rate, either spatially or temporally. • When transporting media over noisy channels packet loss or packet delay can occur. Such losses or delays can affect both the semantics and the syntax of the media stream. The distortion metrics can be categorized into: 1) in terms of the number of measurement results: composite metrics(e.g., PSNR, SSIM) and component metrics(e.g., blocking, blurring, etc.). 2) in terms of measurement region in an image: global and local 3) and others(subjective/objective, FR/NR,…) Conventional image coders mostly rely on composite and global distortion metrics. Since the composite metric can be decomposed into a number of components, new image coders based on component-wise distortion metrics seems to be feasible. By analyzing and reducing any3 component(s) of coding artifacts, it is able to design more efficient encoder. By standardizing the distortion metrics, it may be used for • Transcoding (for existing standards) • Rate-distortion control(for existing and new standards) • Quality requirement for ne w standards Under these background and motivations, we reviewed several No-Reference (NR) and Full-Reference(FR) objective quality metrics and developed some useful algorithms in order to take into account human visual properties and subjective measurement results. 2. Distortion Metrics We have investigated in three types of artifacts, namely blockiness, blurriness and ringing. As already known, the blockiness is the main artifact in block-based coder such as JPEG, MPEG-x and H.26x series, while the blurriness and ringing artifacts occur in the Wavelet-based JPEG2000 encoder. In this section, we reviewed several FR/NR metrics to measure perceptual artifacts in image or video, and presented some modified version of both FR and NR metrics. 2.1. Blocking metrics 2.1.1. Entropy-based blocking metric In general, the blocking artifact is defined as the discontinuities found across block boundary. It is more apparent in smooth textured region where local variance is lower than others [Shao2007]. From these properties, when designing a new metric, we assumed that the blocking dominant region produces lower local variance and entropy than that in original image. We developed the entropy-based metric as shown in Figure 1. First, the artifact regions are extracted by comparing variance and entropy of test and original image. Next, we find the boundary of blocking region based on morphologic operation. Finally, the blocking artifact is measured in a similar fashion of the extrapolation-based blocking metric in [Caviedes2003]. For the NR measurement, we compare local entropy with specified value instead of reference image. This metric shows better measurement results in terms of quality and correlation factor along with perceived distortion (DMOS: Differential Mean Opinion Score) which is depicted in Figure 13. However, threshold value of entropy and variance is a practical problem to decide whether blocking happens or not. The blocking artifacts can be measured as the number of 8-pixel edges found on the 8×8
View Full Document
Unlocking...