No-reference Stereoscopic Image Quality Predictor using Deep Features from Cyclopean Image

With the expanding use of stereoscopic imaging for 3D applications, no-reference perceptual quality evaluation has become important to provide good viewing experience. The effect of the quality distortion is related to the scene’s spatial details. Taking this into account, this paper introduces a blind stereoscopic image quality measurement using synthesized cyclopean image and deep feature extraction. The proposed method is based on Human Visual System (HVS) modeling and quality-aware indicators. First, the cyclopean image is formed, taking on the existence of binocular rivalry / suppression that includes the asymmetric distortion case. Second, the cyclopean image is decomposed into four equivalent parts. Then, four Convolutional Neural Network (CNN) models are deployed to automatically extract quality feature sets. Finally, a feature bank is then created from the four patches and mapped to quality score using a Support Vector Regression (SVR) model. The best known 3D LIVE phase I and phase II databases were used to evaluate the efficiency of our technique. Compared with the state-of-the-art stereoscopic image quality measurement metrics, the proposed method has shown competitive outcomes and achieved good performance.