Hybrid neural systems for reduced-reference image quality assessment

Reduced-reference paradigms are suitable for supporting real-time modeling of perceived quality, since they make use of salient features both from the target image and its original, undistorted version, without requiring the full original information. In this paper a reduced-reference system is proposed, based on a feature-based description of images which encodes relevant information on the changes in luminance distribution brought about by distortions. Such a numerical description feeds a double-layer hybrid neural system: first, the kind of distortion affecting the image is identified by a classifier relying on Support Vector Machines (SVMs); in a second step, the actual quality level of the distorted image is assessed by a dedicated predictor based on Circular Back Propagation (CBP) neural networks, specifically trained to assess image quality for a given artifact. The general validity of the approach is supported by experimental validations based on subjective quality data.