Image Quality Assessment Using a Combination of Hand-Crafted and Deep Features

No-reference image quality assessment is an important area of research and has gained significant interest over the past years. Full-reference image quality assessment is not feasible in some scenarios where the reference image is not available. Hand-crafted features are statistics of natural scene images which are used to train a regression algorithm. Deep learning based approaches learn discriminatory features from images. This paper proposes a hybrid method to Image Quality Assessment (IQA) which uses combination of hand-crafted and deep features. The hand-crafted features are extracted in multi-scale and color space configuration in order to capture greater details. Deep features are extracted using transfer learning of Vgg19. Dimensionality reduction is achieved using principal component analysis and features with 95% variance are retained and a final feature set of 102 transformed features is obtained. Gaussian process regression using the squared exponential kernel is used for modeling. The final model is tested on seven benchmark databases for correlation of estimated image quality and mean of subjective score. A comparison with twelve state-of-the-art methods is performed and superior performance is achieved.