Image Quality Assessment Based on Distortion Identification Aladine Chetouani, Azeddine Beghdadi L2TI, Institut Galilée – University Paris 13, 99 avenue Jean-Baptiste Clément, 93340 Villetaneuse, France ABSTRACT A New Global Full-Reference Image Quality System based on classification and fusion scheme is proposed. It consists of many steps. The first step is devoted to the identification of the type of degradation contained in a given image based a Linear Discriminant Analysis (LDA) classifier using some common Image Quality Metric (IQM) as feature inputs. An IQM per degradation (IQM-D) is then used to estimate the quality of the image. For a given degradation type, the appropriate IQM-D is derived by combining the top three best IQMs using an Artificial Neural Network model. The performance of the proposed scheme is evaluated first in terms of good degradation identification. Then, for each distortion type the image quality estimation is evaluated in terms of good correlation with the subjective judgments using the TID 2008 image database. Keywords: Image Quality Assessment, Artefacts, Classification, Artificial Neural Networks, Subjective Evaluation 1. INTRODUCTION Many Full Reference (FR) Image Quality Metrics (IQMs) have been proposed in the literature. Some are based on pixel- wise differences such as the most common metrics the Peak Signal to Noise Ratio (PSNR) and the Mean Square Error (MSE). Those IQMs are simple and practical for real applications. However, they are not consistent with subjective evaluation. To overcome this limitation, some authors proposed to incorporate into the PSNR some characteristics of the Human Visual System (HVS) such as the Contrast Sensitivity Function (CSF) and the masking effects [1][2][3]. Others IQMs, such as the Visible Difference Predictor (VDP), are fully based on HVS characteristics. As an alternative to fully HVS-based metrics and MSE-like measure are structural-based or mutual information-based such as SSIM [5] and VIF [6] have been proved to achieve a god compromise. However, very often it is implicitly assumed that the developed metrics would be able to quantify any type of degradations. Whereas, the perceptual image quality of a given degraded image depends strongly on the type of the distortion. Indeed, for example a blurring effect has not the same visual impact than blocking or ringing effect on a given image. In this work, a new Global Full Reference Image Quality Metric (FR-IQM)System for estimating the image quality level for many types of distortions is proposed. This system consists of many stages. During the first step, the type of the degradation contained in the image is identified using a Linear Discriminant Analysis (LDA) classifier. Some features are therefore extracted. Then, the degradation-dependent Image Quality Metric (IQM-D) is used to estimate the image quality. The paper is organized as follows: Section 2 presents the limitation of FR measures through a simple experience and the database used in this study. The next section describes the degradation classification process and the corresponding results. Then, the degradation-dependent Image Quality Metric (IQM-D) and its evaluation are presented in the following section. The last section is devoted to conclusion and perspectives. 2. FULL REFERENCE METRICS LIMITATION During the last two decades, many FR-IQMs have been proposed in the literature. However, at present time there is no universal metric that can be used to estimate the image quality. Indeed, all the proposed FR metrics are generally developed without taking into account the type of the degradation which implies that some IQMs provide well results for a given degradation type and obtain poor results for some others. Thus, it would be inappropriate to use the same IQM