No reference image quality assessment for JPEG2000 based on spatial features

Abstract

Perceptual image quality evaluation has become an important issue, due to increasing transmission of multimedia contents over the Internet and 3G mobile networks. Most of the no reference perceptual image quality evaluations traditionally attempted to quantify the predefined artifacts of the coded images. Under the assumption that human visual perception is very sensitive to edge information of an image and any kinds of artifacts create pixel distortion, we propose a new approach for designing a no reference image quality evaluation model for JPEG2000 images in this paper, which uses pixel distortions and edge information. Subjective experiment results on the images are used to train and test the model, which has achieved good quality prediction performance.

Introduction

It is an ever increasing requirement to send more multimedia data over tighter bandwidth which has been driven to develop advanced compression technology. Due to the advanced development of different image compression techniques and processing systems, there is a very big concern about the levels of image quality both for providers and users in many image processing applications from compression to printing. Obviously, digital images suffer a wide variety of distortion in these applications and perceptual quality of the images are degraded. Therefore, perceptual image quality measurement is an important problem. Though the subjective test is considered to be the most accurate method since it reflects human perception, it is time consuming and expensive. Furthermore, it cannot be done in real time. As a result, developing objective image quality evaluation methods are getting more attention nowadays. There are three types of methods that are used for objective image quality evaluation: full-reference (FR), reduced-reference (RR) and no-reference (NR). In the FR method, a reference/original image is required to assess the quality of the distorted image. Therefore, it is highly desirable to develop a quality assessment method that does not require full access to the reference images. In the RR method, some extracted features of the reference/original image are required to assess the quality. However, in many practical applications, the reference image is not available and an NR quality assessment approach is desirable.

The most widely used objective image quality/distortion metrics are peak signal-to-noise ratio (PSNR) and mean squared error (MSE), but they are widely criticized, among others things, for not correlation well with perceived quality measurement. In the past, a great deal of effort has been made to develop new objective image/video quality metrics that incorporate perceptual quality measures by considering human visual system (HVS) characteristics [1], [2], [3], [4], [5], [6]. Most of the proposed image quality assessment approaches require the original image as a reference.

Nevertheless, human beings do not need to have access to the reference image to make judgements regarding quality. Human observers can easily assess the quality of distorted images without using any reference image. By contrast, designing objective NR quality measurement algorithms is a very difficult task. This is mainly due to the limited understanding of the HVS, and it is believed that effective NR quality assessment is feasible only when prior knowledge about the image distortion types is available. Although only a limited number of methods have been proposed in the literatures for objective NR quality assessment, this topic has attracted a great deal of attention recently [7], [8], [9], [10], [11], [12], [13], [14], [15], [16], [17], [18], [19]. Since the predominant mode for image and video coding and transmission is using block-based video compression algorithms, blind measurement of the blocking artifact has been the main emphasis of NR quality assessment researches [7], [8], [9], [10], [11], [12], [13]. Blockiness, activity and segmentation-based measure are explained either in the spatial domain [7], [8], [9], [10], [11] or in the frequency domain [12], [13].

However, the above described methods would obviously fail for any other distortion types, such as ringing or blurring introduced by the JPEG2000 image compression algorithm, or the H.264 video compression algorithm. Some researchers have attempted to quantify the blurring and ringing artifacts without reference. In [14], a visible ringing measure (VRM) is proposed that captures the ringing artifact around strong edges. The algorithm is based on constructing an image mask that exposes only those parts of the image that are in the vicinity of strong edges, and the ringing measure is considered to be the pixel intensity variance around the edges in the masked image. Still, the measure was not compared to the human score of quality. In [15], [16] an NR blur metric is proposed based on measuring average edge transition widths, and this blur measure was used to predict the quality of JPEG2000 compressed images. In [17], an NR algorithm is proposed based on natural scene statistics. In [18], a principal component analysis is performed on edge points, beforehand classified as distorted or not, in order to measure both blurring and ringing effects, as well as the combination of spatial ringing and blurring measures, which are also presented in [19].

All of the proposed NR perceptual image quality assessment algorithms are implemented according to the predefined specific artifacts of specific coders [7], [8], [9], [10], [11], [12], [13], [14], [15], [16], [17], [18], [19]. A lot of NR quality evaluations have already focus on measuring the blocking artifacts, specially for JPEG images with quite sufficient correlation with the subjective scores. However, very few NR evaluations have been performed for JPEG2000 images and these evaluations’ generalization ability and performances are not widely verified and well matched to the subjective scores. Whereas, nowadays, the JPEG2000 coder is getting more attention compared to the JPEG coder, due to its high coding performance, although JPEG was previously the standard coder for still image. The JPEG2000 coder served better in many image processing applications, such as digital cameras, 3G mobile phones, video streaming, printers, scanners, high quality frame-based video recording, nonlinear video editing, storage, etc. Specifically, motion JPEG2000 is the leading digital film standard currently supported by digital cinema initiatives for the storage, distribution and exhibition of motion pictures.

In this research, we propose a new method for NR quality evaluation of JPEG2000 images, irrespective of any predefined specific artifacts based on pixel distortions and edge information measure. This type of quality assessment is used to assess image quality and produces results comparable to those of subjective scores. The subjective experiment results on our database (JPEG2000 color images) were used to train and test the model and it achieved a sufficient quality prediction performance. The other database was also used to verify the model's performance. We report that the performance of the model is sufficient and reliable.