A hybrid system for distortion classification and image quality evaluation

Numerous Image Quality Measures (IQMs) have been proposed in the literature with different degrees of success. While some IQMs are more efficient for particular artifacts, they are inefficient for others. The researchers in this field agree that there is no universal IQM which can efficiently estimate image quality across all degradations. In this paper, we overcome this limitation by proposing a new approach based on a degradation classification scheme allowing the selection of the ''most appropriate'' IQM for each type of degradation. To achieve this, each degradation type is considered here as a particular class and the problem is then formulated as a pattern recognition task. The classification of different degradations is performed using simple Linear Discriminant Analysis (LDA). The proposed system is developed to cover a very large set of possible degradations commonly found in practical applications. The proposed method is evaluated in terms of recognition accuracy of degradation type and overall image quality assessment with excellent results compared to traditional approaches. An improvement of around 15% (in terms of correlation with subjective measures) is achieved across different databases.