Research On Texture Segmentation And Texture Replacement In Image

Texture is a popular research topic in computer graphics, computervision and image processing areas. In the last three decades, as one of themost significant and challenging research perspective, texture analysis hasattracted particular attention from researchers, and resulted in a series ofresearch achievements in this topic. Among these achievements, there are alot of texture analysis methodologies such as Coexistent Matrix method,Markov Model approach, wavelet transformation method, Tamura Modelapproach and to name a few. Using texture analysis, textural features can bequantitatively analyzed and specified, and then can be applied into researchstudies in texture segmentation, texture synthesis, and texture substitute.There have been numerous methods in texture segmentation. Generallyall texture segmentation methods can be classified into four categories, theyare Regression Models based on statistical methods;structure-based methodssuch as Voronoi diagrams;model-based approaches such as fractal model;approaches that are based on dimension frequency analysis / multi channelfiltering methods, for instance wavelet approach, and Gabor filteringapproach. According to reports in biological studies, the visual system ofhuman beings can be deemed as a multiple-channel filtering model. Becausedimension frequency analysis / multi-channel filtering approaches reflectthese properties in human's vision systems, more and more attention andapplication are applied to these types of approaches.This article proposed a DEMD and sampling based texture segmentationalgorithm, which can extract the texture sampled by user in image. Accordingto the taxonomy in texture specification, it is a type of feature analysissegmentation method based on dimension frequency. In this algorithm,textures are analyzed using DEMD (Direction Empirical ModelDecomposition), then in the inherent texture direction, the 2D textural imagecan be decomposed into a series of stable IMF signals and a residual signal.Texture signature can be obtained by extracting frequency and closure fromthese stable IMF signals. Because this method has addressed the inherenttexture direction and stableness of the IMF signals, it can generate moreaccurate and representative texture signature.Empirical Mode Decomposition method was proposed by Huang et al in1998. EMD approach first decomposes signals using "screenings", thenemploys Hilbert transformation to define the compound signal and instantfrequency for each element. Based on one-dimensional EMD, Liu et alpresented two-dimensional EMD method to extract textural signatures ofimages. This method has full data-driven self-adaptiveness and can rendertextural signatures which reflect unique visual meaning. Hence this methodopens a new horizon for texture segmentation research. Previously, texturalsignatures of Gabor filtering and Wavelet approaches are actually closuresgenerated by space, frequency/length locality filters. Since frequencies havebeen pre-defined by the filter frequency, thus it cannot be used as signature;on the other hand, it is also problematic to use the instant frequency of theoriginal signal as the texture signature. For example, multiple-factoredsignals only have a single frequency at one particular moment, henceforth atsome points the frequencies might be negative in this regard. EMDapproaches elegantly solves the problems mentioned above.In addition, the texture segmentation algorithm presented in this thesisdetermines the texture pattern to be extracted according to user-selectedtexture sample. Through the texture sample, the algorithm can determinewhether each pixel in the image belongs to inner area or external area of theextracted texture. It well addresses the over-segmentation problem oftenoccurring due to texture local variations.The DEMD and sample based texture segmentation method in the thesiscan be considered as a global optimization problem. It segments texturessimulating random degrading methods, and renders satisfactory results. Thesegmentation candidate areas can be obtained using correlation matrixmethod, which confines texture segmentation to a certain area, and reducesthe search space for simulating degrading, thus dramatically increases thespeed of texture segmentation.Texture synthesis is proposed to address the distortion problem intexture mapping. The goal is to generate a similar texture image but not anexact copy as the sample texture when it limited and known. In 2001, Efrosand Liang simultaneously proposed a MRF-based model and patchcombination texture synthesis approach which renders high-quality synthesisby exploiting adjacency correlation between texture patches. This approachhas several advantages such as high speed, high quality and adaptive to widerange of textures, hence it boosts synthesis techniques to a new high, andleads to the future direction of synthesis approaches.All previous patch based texture synthesis methods directly usebrightness value or RGB value of sample texture to select matching patchesfor synthesis through error calculation. However, for some textures especiallythose with relatively large base and higher randomness, the obtained patchthrough boundary matching may not be ideal, due to lack of consideration intexture multi-dimension when directly using brightness value and/or RGBvalues of sample texture for calculation. Although the overall matching errorof the combined texture patches is still acceptable, in fact the texture patchesdo not match the same in each texture dimension, and the visual impact ofdifferent texture dimensions are different. For example, a bigger matchingerror in a stronger-structured decomposed component will result in aunsmooth seaming. Additionally, extremely small matching error restraintwill result in an exact copy of the texture sample, so existing patch-basedtexture synthesis usually relax on matching error constraints in order toincrease the number of candidate matching texture patches. However, suchrelaxation will lead to unsmooth seaming at the border of patches in thesynthesis image. This kind of unsmooth seaming cannot be completedremoved, although some methods such as eclosion can alleviate the seaming.This thesis proposed a EMD based multi-dimensional texture synthesisapproach. Through 2D EMD, the texture sample can be decomposed intomultiple IMF images and a residual image;then patching and synthesis canbe conducted at various dimension using the correlations among the variouslevels of images;in the end the final image can be generated by integratingimages in multiple dimensions. The synthesis approach proposed in thisarticle took into account multiple-dimension feature of a texture, andgenerates synthesis in different dimensions. Compared with patch-basedapproaches, our algorithm can generate more natural images and retain themacro-randomness of texture.Texture can be described as basic pattern repetition in different spatialdirections, it is a cellular organic phenomenon. Therefore, for strongstructured textures, structure-based synthesis approaches are normally used.In these approaches, cell shape and properties need to be determined, andthen the position of the cells need to be confirmed. Near-regular textures arethe most representative structured textures, and can be easily found in thenature. This type of textures are those regular textures that deviate regularlyin different dimensions.This thesis introduced the concept and taxonomy of near-regulartextures, and studied grid-adjustment texture analysis approach. Thisapproach is mainly used in near-regular texture analysis, using user-adjustinggrid to represent texture structure. It analyzes textures based on second-classtransformation field definitions, then calculates geometric transformationfield and brightness coefficient to get near-regular textures geometricstructure distortion function and brightness coefficient distribution function.Texture replacement is one of the applications of texture research. Itrefers to the techniques to maintain unchanged light, shading, and physicalstructure distortion while replacing a particular texture pattern. It has broadrange of usages in real life, for instance interior design, digital movie, virtualreality and computer vision, etc.For a particular texture in the image, its visual effect in the image isdependent on three factors: its inherent texture pattern, external lightcondition, and texture's physical structure distortion due to its spatial position.The purpose of texture replacement is to replace a particular texture patternwith another texture, meanwhile to maintain the same external light conditionand physical structure distortion. In other words, it changes the first condition(inherent texture pattern) and retains the rest two to render a new visual effectunder the same light condition and geometric structure distortion.This article studied texture replacement problem, proposed its owntexture segmentation based on DEMD and sample-based texture, and its owntexture synthesis approach based on EMD multi-dimension texture synthesisalgorithm and grid-adjustment texture synthesis method. The experimentalresults of the algorithms are satisfactory and promising.