Optimum Framework Design for No Reference Image Quality Assessment based on Generic Statistical Modeling

Abstract

Since its inception, a digital image undergoes various image processing tasks and applications where different distortions are introduced in an image. These distortions are required to be quantified for quality estimation. In the real-world applications, image quality has to be estimated in the absence of the pristine image. No Reference Image Quality Assessment (NRIQA) methods play a vital role in such situations. The digital images may have suffered from different distortions. No single image feature can model the presence of the distortions in an image. Therefore, a generic approach of combining parametric and other level-two features is proposed. This hypothesis is tested with the combination of existing block based, second-order statistical, and natural scene statistics based features. In the subsequent experimentations textural, block based, and shape adaptive wavelet transform features are implemented. The experiment further extended to derive a feature vector through a fusion of scale invariant feature transform key points and curvelet coefficients. Spatial covariance is also implemented as a feature. This study proposes a framework for NRIQA to estimate the overall image quality. In order to implement the framework that can address different distortions, it is trained on the feature vector created by combining the above features. A combined feature vector grows in size with every appended feature and may raise the dimensionality curse problem. Therefore, to better the efficiency, the combined feature vector is optimized with principal component analysis. The relationship between extracted features, subjective score, and the predicted score is complex and non-linear. Therefore, a machine learning based approach is proposed for designing the framework. SVM and ANN are employed as the quality prediction tools in the initial experimentations. It has been observed that the artificial neural network remains undertrained when it is fed with the feature values with lesser variations, and the feature magnitudes are not proportional to the degradation. Therefore, the neuro-wavelet model is developed. Neurofuzzy classifier is also implemented. The performance of the metrics are evaluated through LCC and SROCC. NFC and Neuro-wavelet model shows better accuracy for optimized combined feature metric. The proposed framework is developed to work with gray images and tested on LIVE, TID2008, and TID2013 datasets.