No-Reference Stereoscopic Image Quality Algorithm Based on Features on DCT Domain

Tian Liu1, Yuxia Sheng1, Li Chai1, Jingxin Zhang2

  • 1Wuhan University of Science and Technology
  • 2Swinburne University of technology

Details

10:50 - 11:10 | Mon 19 Aug | Lau, 6-211 | MoA4.2

Session: Learning Applications

Abstract

Existing image quality assessment methods can not extract stereoscopic features effectively. In this paper, a no-reference quality assessment method is proposed for stereoscopic images based on DCT domain features. The main steps of the proposed methods areas follows: (1)Synthesize the left and right images to two “cyclopean”maps respectively according to the binocular fusion characteristics and the binocular rivalry characteristics. (2)Extract the statistical features in DCT domain by using normalized DCT coefficient of the distorted image pairs and the obtained two “cyclopean”maps. (3)Use support vector regression(SVR) to predict the objective scores of stereoscopic images by establishing a map between the stereoscopic image feature vector and the difference mean opinion score(DMOS). The proposed method is evaluated by images from the LIVE Phase I and the LIVE Phase II database. Experimental results show that compared with other methods, the proposed method is consistent with subjective perception, and hence outperforms the existing methods.