Research On Temporal-Spatial Video Modeling And Denoising

As video signal is sampled continuously from objective world, strong relativity exists between adjacent pixels, which could promote video signal processing. Spatial-Temporal is imported to describe the complex relativity of space-domain and time-domain for video semantics analysis and video processing promoting. Temporal-Spatial method has become an important way to discover potential information in videos. Separating the relativity in space-domain and the relativity in time-domain, conventional video models view video signal as random variables. So with CVM video processing can but be performed separately in space-domain and time-domain, and Temporal-Spatial Processing is performed on the results. Video models express the spatial and temporal characters of video must be developed for more thorough Temporal-Spatial Processing. Problems about Temporal-Spatial Processing, Temporal-Spatial video modeling, Temporal-Spatial video denoising, Temporal-Spatial video object detecting and tracking, is studied in this dissertation, and several achievements have been obtained.l.A Video Grid Model (VGM) is proposed. VGM decompose videos with Video Grids, which have only 3 movements, translation, rotation and scale. VGM expresses time-domain relations with Video Grid's movements, and expresses space-domain relations with Video Grid's area and layer. Five steps, Region Segmenting, Region Matching, Layers Determination, Motion Estimation, and Region Resegmenting, are involved in VGM obtaining. The original Video Grid regions can be divided by colors and matched with semantics of edges. The region layers can be decided by shading. The Motion Estimation between regions can be performed via a least-square estimation with inflexions on edges. Regions are resegmented when the error is large, until the size Sum of Abstract Difference of matched pixels is allowed.2.A Video Flow Model (VFM) is proposed. The VFM view all pixels from a same object area as a flow. Video Flow Line (VFL) is employed to express the center motion. Video Flow Trace (VFT) is employed to express the point motion. The VFL is smooth in time-domain, and the VFT is smooth in both time-domain and space-domain. The VFT can be simplified by Cut off Line, Interpolation Line, Rotation Operator, and Translation Operator. Three steps, Region Segmenting, VFL computation, and VFT computation, are involved in VFM obtaining. The original regions of Video Flow can be divided by colors. The VFL can be predicted by the smoothness in time-domain and modified by centroid of regions. The VFT can be computed by invariance of pixels on same VFT, smoothness in space-domain, and smoothness in time-domain. The VFT can also be estimated by isochromate. The Video Flow regions can be resegmented and combined according to the spatial smoothness of VFT.3.Video denoising based Temporal-Spatial video models is studied. VGM based Median video Denoising method decides neighbor by Temporal-Spatial relations in VGM. VGM based Median video Denoising method determines reference value of pixels by pollution degree. It is suitable for removing impulse noise of high density. VFM based Median video Denoising method determinates the reference value with smoothness of VFT in time-domain and space-domain. VGM based Interpolation video Denoising method restores smoothness of Video Grid regions by interpolation. VFM based Interpolation video Denoising method restores smoothness of VFT by interpolation. Experiments shows that video denoising based Temporal-Spatial video models stand out of pure time-domain methods and space-domain methods. It surpasses methods without Temporal-Spatial video model too.4. A Temporal-Spatial video denoising method based Motion Compensation is proposed. In this method Kalman filter is employed for temporal denoising and median filter is employed for spatial denoising. The reliablity is computed with pollution degree and used in computation of geometric mean of temporal denoising result and spatial denoising result, which is treated as the final denoising results. The Kalman filter and median filter based method surpassed the Temporal-Spatial denoising methods to remove impulse noise in the past.5. The Temporal-Spatial video models are applied in video monitor of steel-tube production. An object detecting and tracking method based VFM is proposed. It's easy to recognize humans, steel-tubes, devices, and ground, as their VFT features is easy to obtain in VFM of steel-tube produce videos. In steel-tube object tracking based on VFM, the object can be predicted by VFL and VFT with their temporal smoothness and modified by the spatial smoothness of VFT. Mechanism of motion tracking and combine with multi-camera is also built up in this dissertation to meet the production needs. Experiments show that target tracking based VFM has small error and high speed, so it is suitable for timely and precise tracking.