Automatic Distortion Type Recognition for Stereoscopic Images

Abstract

Stereoscopic image quality evaluation and enhancement are facing more challenges than its 2D counterparts. The use of stereoscopic/3D imaging is rapidly increasing. Stereo images could be afflicted by different types of distortion. For the development of stereoscopic image quality evaluation and enhancement algorithms, a no-reference distortion classification model has been proposed. Disparity/depth map is constructed and utilized as a source of information. Gradient map variance is extracted as feature from disparity and stereo image. Following feature extraction, a machine learning based on Support Vector Machine (SVM) has been employed to learn and identify the distortion type. The model is trained and used to classify the most common types of distortions. The benchmark database LIVE 3D has been used to test and evaluate the model. Testing results of the proposed classifier have shown reliability and good accuracy on five types of distortions.