Key Technology Research Of Bilateral Filter

Bilateral filter has appeared in the 1990s,for it has simple,noniterative and intuitive formulation and excellent edge-aware smoothing ability,many researches are focused on its acceleration,improvement and application.For more than 20 years,more and more important problems in computer vision and graphics can be perfectly solved through bilateral filter,which also promotes the modification and improvement of bilateral filter.With the progress of data acquisition devices,the resolution and depth of images and videos are dynamically increasing,therefore,puts forward more strictly requirements on the runtime and performance of bilateral filtering algorithm.To solve the existing key problems of bilateral filter,the main content of the disseration is:1.In view of the runtime of bilateral filtering which depends on the size of local window,cannot meet the requirements of parctical application,we utilize two range ker-nel approximation methods,and achieves accurate acceleration algorithms which reduce the compuational complexity of bilateral filter to linear complexity.These two methods are based on the exponential sum approximation of the range kernel of BF,we can decom-pose the original BF into a bunch of spatial filters of which the range kernel is shiftable ex-ponential function.This kind of spatial filter can be computed in linear time,independent of the size of local window,therefore,their linear combination,the well-approximated BF can also be executed in linear time.The two methods adopt different ways to approximate the range kernel.In the first acceleration method,an exponential sum is exploited to ap-proximate the range kernel of the bilateral filter,where the coefficients of the exponential basis are computed by solving a set of linear equations.In the second method,we first obtain the Z transformation of the range kernel,then approximate the Z transformation of the range kernel using the Pad?eApproximation.Finally,we inverse the transformation of the Pad?eApproximation and obtain an exponential sum to approximate original range kernel,where the coefficients of the exponential basis are computed by solving a set of linear equations.Experiments show our two methods achieve state-of-the-art results in terms of accuracy and speed.2.In view of the edge-preserving ability of bilateral filter is not satisfactory which will blur the edges when remove noises,we combine the local filtering with weighted l₀ nonlocal total variation?NTV?optimization,and achieve l₀ bilateral filter which own-s enhanced edge-preserving ability.Although bilateral filter is edge-aware filter,it will inevitably blur the edges when filtering the image.To improve the performance of BF,we design a new scheme to combine the weighted l₀ nonlocal total variation?NTV?opti-mization with the original BF named l₀ bilateral filter.The l₀ bilateral filter which derived from the weighted l₀ nonlocal total variation?NTV?optimization automatically inherits the excellent ability of edge-preserving.On one hand,in image smoothing,l₀ bilateral filter preserves the sharp edges and suppresses the different scales noise simultaneously.On the other hand,in edge enhancement and extraction,it can restore high quality edges in complex environment and inevitable noise.3.In view of smoothing out high-contrast details while preserving major structures in not achievable for bilateral filter,we utilize the bilateral weight computation scheme based on tree distance defined on segement graph,and achieve segment graph based bi-lateral filter which can smooth out the high constrast detail.Bilateral ilter determines the pixels'similarity throught the postion and color/intensity differences,this kinds of weight computation scheme cannot smooth out high-contrast details.To endow BF the ability of removing high-contrast details and textures,the connective similarity basd on tree distance of segment graph is used to replace the spatial weight with the connectivity weight.The newly designed segment graph based bilateral filter can also avoid the”leak”artifact and can be computed in linear time.The experiments in high-contrast detail-s/textures smoothing,scene simplification,edge/boundary extraction and texture editing demonstrate the superiority of SGBF.4.In view of the needs to accurately recogonize the tangrams in the video in real time,we use a tangram recogonition method based on fast bilateral filtering,and achieve fast and accurate recognition results.BF is usually utilized in video denosing and en-hancement,and the tangram recognition needs to preprocess the video frame to remove the noise and sharpen the edges.Most of exsisting edge-aware filters have large compu-tation cost or the filtering result is unsatisfactory,thus are unable to meet the needs of real time processing requirements,so we integrate the accelerated bilateral filtering algorithm into the video tangram recognition problem.Firstly,we use the accelerated BF to denoise the video frames,then empoly the GMM model to perform the background subtraction.Secondly,the foreground pixels are classified into seven classes corresponding to the d-ifferent tangrams.Finally,the initial recognition results are refined through morphology operations and polygon approximation to get the location and outline of tangrams.The experiments show that our method can accurately and stablely recongnize the tangram in real time.