| 英文摘要 |
Video magnification technique can adjust the amplitude of small changes in the video by processing the spatio-temporal information. This technique can enhance the changes with low amplitude in the world for perception and visualization, which are hard to be perceived by the naked eye. Furthermore, it can reveal the important information and laws contained in the small changes in the world, and can help people better perceive and understand the small changes in the dynamic world through image sequences. In summary, the study of the video magnification technique is of great research value and practical significance. Eulerian video magnification (EVM) provides an effective approach to amplify the changes with low amplitude, which deal with the small changes in the video based on Eulerian perspective. The small changes are extracted and magnified by the spatiotemporal information processing. Unfortunately, the current methods based on Eulerian perspective have some shortcomings and problems that are sensitive to noise and susceptible to large motion interference. The noise in video increases with the increase of magnification factor. When the video is in presence of large motion in video, the magnification results will produce serious blurring and significant artifacts. As to these problems, this thesis presents an improved video magnification algorithm with anti-noise performance and two advanced algorithms with anti-interference of large movement. The study works of this thesis are elaborated as follows: First, aiming at the problem that Eulerian video amplification based on brightness changes is sensitive to noise, this thesis proposes a new video magnification algorithm based on principal component analysis (PCA). The algorithm firstly analyses the difference between temporal and spatial variations of video content and noise characteristics based on the Eulerian perspective. Then the small variations in the image sequence are extracted by the principal component analysis which can suppress noise interference in the processing. Furthermore, an improved video magnification algorithm with anti-noise performance is studied and established. Finally, the experimental results verify the effectiveness of the presented algorithm. Second, considering the problem that Eulerian video amplification method is disturbed by non-linear large motion, a new anti-interference of large movement video magnification algorithm is proposed based on the amplitude selective filtering. In Eulerian framework, we firstly analyze the energy characteristic difference between large motion and small changes in image sequence. Then the small changes are extracted by the amplitude selective filtering which can remove the large motion and maintain the quality of the small. Furthermore, combining with variations characteristics of small change in video, an improved video magnification algorithm with anti-interference of large movement is studied and established. Finally, we provide quantitative as well as qualitative evidence for the presented algorithm while comparing to the state-of-the-art. Experimental results verify the effectiveness of the presented algorithm. Third, because of the problem that Eulerian video amplification method is disturbed by fast large motion, we propose a new video magnification algorithm based on the thirdorder derivative of the Gaussian filtering. The algorithm firstly analyses the difference of acceleration characteristics between fast large motion and small motion in image sequence. Then we extract the small changes in the video against the fast large motion interference by the third-order derivative of the Gaussian filtering. On this basis, the intensity and phase changes characteristics of small motion in image sequence are considered and a novel video magnification algorithm with fast large motion interference resistance is studied and established. Qualitative and quantitative evaluation are performed on real videos as well as on synthetically-generated sequences with ground truth available. Results show that the proposed algorithm can handle large motions well and reduce the artifacts and blurring comparing to the other methods. Last but not least, in consideration of the advantages of Eulerian video amplification technique in the signal with low amplitude processing and sub-pixel precision change signal processing, we develop and design a visual detection system for small changes in the world. Based on the Eulerian video magnification algorithms, different small change signals in specific scenes are detected and revealed, such as the non-contact heart rate (HR) estimation based on video small color change amplification; the non-contact respiratory rate (RR) estimation, the non-contact pulse wave measurement and target micro-vibration measurement based on video small motion magnification. The system can carry out the small-amplitude changes in different scenes measurement and visualization in real-time. KEYWORDS: Eulerian video magnification (EVM); PVM; principal component analysis (PCA); Gaussian derivative filtering; heart rate (HR) estimation; respiratory rate (RR) estimation; pulse detection; vibration detection
|
