TY - JOUR
T1 - 基于运动路径全曲率的视频稳定质量评价
AU - Zheng, Qing Zhuo
AU - Zhang, Lei
AU - Huang, Hua
N1 - Publisher Copyright:
© 2018, Science Press. All right reserved.
PY - 2018/11/1
Y1 - 2018/11/1
N2 - A video captured with a hand-held device (e.g., cell-phone or tablet computer) often appears remarkably shaky. It has a great negative impact on people's visual perception and is difficult for the following processing of the video. As a technology that can remove video jitter to enhance visual quality, video stabilization algorithms have been widely and deeply studied in the last ten years. However, how to effectively evaluate the performance of video stabilization algorithms is a topic that has not been studied deeply. Due to the lack of a comprehensive and fair evaluation mechanism, it is impossible to judge the performance of video stabilization algorithm objectively, which makes it difficult for users to choose video stabilization algorithm. The video stabilization objective assessment can be divided into two ways: non-reference and full-reference. The full-reference method needs the benchmark video, then calculates the difference between the evaluated video and the benchmark video, and then evaluates the stability of the video. Non-reference method is used to analyze the jitter characteristics of video directly through the motion information of video, and then evaluates the video stability. Compared with the full-reference methods, non-reference methods are more flexible and economical. But the existing non-reference video quality assessment algorithms have some problems, such as sensitive to video with severe jitter or inconsistent with the subjective assessment. In view of these problems, a non-reference video stability assessment method for directly measuring the smoothness of video motion path is proposed. Firstly, we detect the feature points of adjacent frames, and then calculate the inter-frame homography matrix. Homography matrix can accurately describe the translation, rotation and other transformations between video frames. Secondly, we map the homography matrices in the Lie Group space to form the motion path. Finally, we calculate the total curvature by using the discrete geodesic approximation method to measure the smoothness of the motion path, and use it to evaluate the stability of the video. In order to verify the efficacy of the evaluation, we construct a dataset of 150 video clips and their corresponding stabilized videos based on the publicly available dataset. Stabilized videos in this dataset are the results of five video stabilization algorithms. We choose 2 classic stabilization algorithms from published literature: the spatially and temporally optimized method (STO) and bundled paths methods (BP). Besides, we choose three sophisticated softwares that realise the stabilizer function: Adobe After Effects (AE) warp stabilizer, Google YouTube stabilizer and VirtualDub Deshaker. The experimental result shows that the correlation between the results obtained by the proposed algorithm and the subjective evaluation results is 97%. We make comparisons with some other classic video stabilization quality assessment methods based on this dataset. We collect two representative methods that are used in previous video stabilization methods: inter-frame transformation fidelity (ITF) and low/high-frequency rate (LHR). Compared with the evaluation algorithm based on ITF, the correlation improves 39%. Compared with the evaluation algorithm based on LHR, the correlation improves 21%. The consistency with the full-reference video stabilization quality evaluation algorithm further verifies the effectiveness of the proposed algorithm.
AB - A video captured with a hand-held device (e.g., cell-phone or tablet computer) often appears remarkably shaky. It has a great negative impact on people's visual perception and is difficult for the following processing of the video. As a technology that can remove video jitter to enhance visual quality, video stabilization algorithms have been widely and deeply studied in the last ten years. However, how to effectively evaluate the performance of video stabilization algorithms is a topic that has not been studied deeply. Due to the lack of a comprehensive and fair evaluation mechanism, it is impossible to judge the performance of video stabilization algorithm objectively, which makes it difficult for users to choose video stabilization algorithm. The video stabilization objective assessment can be divided into two ways: non-reference and full-reference. The full-reference method needs the benchmark video, then calculates the difference between the evaluated video and the benchmark video, and then evaluates the stability of the video. Non-reference method is used to analyze the jitter characteristics of video directly through the motion information of video, and then evaluates the video stability. Compared with the full-reference methods, non-reference methods are more flexible and economical. But the existing non-reference video quality assessment algorithms have some problems, such as sensitive to video with severe jitter or inconsistent with the subjective assessment. In view of these problems, a non-reference video stability assessment method for directly measuring the smoothness of video motion path is proposed. Firstly, we detect the feature points of adjacent frames, and then calculate the inter-frame homography matrix. Homography matrix can accurately describe the translation, rotation and other transformations between video frames. Secondly, we map the homography matrices in the Lie Group space to form the motion path. Finally, we calculate the total curvature by using the discrete geodesic approximation method to measure the smoothness of the motion path, and use it to evaluate the stability of the video. In order to verify the efficacy of the evaluation, we construct a dataset of 150 video clips and their corresponding stabilized videos based on the publicly available dataset. Stabilized videos in this dataset are the results of five video stabilization algorithms. We choose 2 classic stabilization algorithms from published literature: the spatially and temporally optimized method (STO) and bundled paths methods (BP). Besides, we choose three sophisticated softwares that realise the stabilizer function: Adobe After Effects (AE) warp stabilizer, Google YouTube stabilizer and VirtualDub Deshaker. The experimental result shows that the correlation between the results obtained by the proposed algorithm and the subjective evaluation results is 97%. We make comparisons with some other classic video stabilization quality assessment methods based on this dataset. We collect two representative methods that are used in previous video stabilization methods: inter-frame transformation fidelity (ITF) and low/high-frequency rate (LHR). Compared with the evaluation algorithm based on ITF, the correlation improves 39%. Compared with the evaluation algorithm based on LHR, the correlation improves 21%. The consistency with the full-reference video stabilization quality evaluation algorithm further verifies the effectiveness of the proposed algorithm.
KW - Homography
KW - Lie Group
KW - Motion path
KW - Total curvature
KW - Video stabilization quality assessment
KW - Visual motion perception
UR - http://www.scopus.com/inward/record.url?scp=85059555140&partnerID=8YFLogxK
U2 - 10.11897/SP.J.1016.2018.02524
DO - 10.11897/SP.J.1016.2018.02524
M3 - 文章
AN - SCOPUS:85059555140
SN - 0254-4164
VL - 41
SP - 2524
EP - 2535
JO - Jisuanji Xuebao/Chinese Journal of Computers
JF - Jisuanji Xuebao/Chinese Journal of Computers
IS - 11
ER -