Research On Stereo Video Objective Quality Assessment Based On Visual Perception

Stereo video can bring a better watching experience for viewer by providing depth information. Thus it attracts more and more focus and becomes a hot topic in study of communication gradually. Meantime, stereo video quality assessment(SVQA) is one of the key technologies in stereo video communication systems. In this thesis, by exploring the characteristics of both human vision and stereo videos, two objective SVQA methods are proposed. The work mainly consists of four parts as described below.(1) As the fundamental part of stereo video, image quality contributes a lot in stereo video quality. By studying the procedure of fractal coding theory, a self-similarity based reduced-reference image quality assessment(IQA) method is proposed. This method utilizes collage error, which is variety used in fractal coding, as image feature. Block quality is calculated by measuring the difference between reference image and distorted one. Eventually, block qualities are combined as image quality according to their saliency. To testify the proposed method on LIVE database, the Linear correlation coefficients(LCC) and Spearman rank order correlation coefficients(SROCC) are above 0.92 and 0.9 respectively. Experimental results demonstrate that the proposed method can predict image quality accurately.(2) Aiming at the difficulty in describing temporal information of video, the thesis proposes a 3D-DWT based SVQA method, which describes temporal information in wavelet domain. Firstly, the proposed method exploits dual-tree complex wavelet transform(DT-CWT) to combine two views’ videos as binocular fusion video(BFV). Then 3D-DWT is used to decompose the groups of frames(Go F) in BFV and calculate each Go F’s quality. Finally, all Go Fs’ qualities are weighted adding as BFV quality, which is stereo video quality. LCC between predict quality and subjective quality is above 0.92, and SROCC is above 0.91, and RMSE is about 5. The proposed method is accordance with human vision.(3) Due to the constantly increasing in resolution of video and image, data quantity in communication systems is tremendously huge. Meanwhile, information of reference image is unavailable in most systems. As a result, the thesis proposed a no-reference IQA method based on distribution feature of wavelet coefficients. Since wavelet coefficients transformed from natural scene image approximate to follow a-stable distribution, parameters of a-stable distribution’s probability density function(PDF) can be extracted as distribution feature. Meantime, distribution feature vector is built between different sub-bands. Finally, prediction model is constructed by support vector regression(SVR). The LCCs and SROCCs between prediction quality and DMOS are all above 0.9, and RMSEs are less than 6.5. The performance index reveals that the proposed method can predict image quality precisely without any information from reference stereo video, which is worth application.(4) Since wavelet coefficients of video share the same distribution pattern with image’s wavelet coefficients, considering the characteristics of stereo video, the thesis proposes a no-reference stereo video quality assessment method based on distribution feature of wavelet coefficients. The proposed method utilizes three-dimensional discrete wavelet transform(3D-DWT) to obtain wavelet coefficients of important Go Fs. Meanwhile, both left and right views’ distribution features are extracted to build prediction model in SVR. The proposed method’s LCC and SROCC in NAMA3 D database are all above 0.85, and RMSE are under 0.65. Experimental results demonstrate that the proposed method performs well in predicting stereo video quality.