Video Quality Assessment Based On Spatio-temporal Transform

Recent years, with the rapid development of the digital video technology and advanced computer network technology, a variety of multimedia technologies have been widely used in people’s daily lives in form of digital camera, video on demand, video conference and video surveillance. However, the quality of videos will inevitably suffer distortions due to the acquisition, storage, processing and transmission. Therefore, it’s of great practical significance to evaluate the video quality.Video signal is quite different from image signal beacause of the time domain. Therefore, it’s more complicated and difficult to be described. So it’s not suitable to simply perform video features extraction and description followed by a two-dimension(2D) transform. In this paper, a video sequence is generally visualized as a volume of data, and can be conceived as a three-dimension(3D) signal rather than a 2D signal.3D spatio-temporal transform is introduced to extract video spatio-temporal frequency information, and design the perceptive features that can accurately describe the video degradation.A reduced-reference(RR) video quality assessment(VQA) method based on three-dimension discrete cosine transform (3D-DCT) has been proposed in this paper. We explore and exploit the statistical characteristics in the 3D-DCT domain and three types of statistical descriptors are designed. Based on an analysis of the relationship between these descriptors and video distortions, we utilize the descriptors computed from both the reference and distorted videos to predict the quality score of distorted videos. Experimental results on the LIVE database demonstrate that the proposed RR-VQA metric correlates with human judgments of video quality quite well, and competes favorably with state-of-the-art RR-VQA methods and even some full-reference(FR) VQA methods.In this paper, we also present a FR-VQA method based on 3D gradient similarity.3D gradient kernels are utlized to video sequence for 3D convolution. In addition, we exploit the local 3D gradient in three directions of reference and distorted videos to design the 3D gradient similarty as the local video quality index. According to the characteristics of human visual system(HVS), the video quality is obtained by spatio-temporal pooling. Experimental results show that the proposed metric can reflect the subjective perception to video quality accurately.