The Research On The Technique Of Texture Synthesis With Its Applications

Texture is widely-used in computer vision and virtual reality and it has importantpractical value. Among the texture-related researches, texture synthesis techniqueswhich address the distortion in texture mapping are most extensively used. Texturesynthesis is to generate a similar textural image based on a limited number of givensample texture images, and the newly generated texture image cannot be an exactcopycat of the original texture image. After decades of research on this topic, someresults have been achieved in recent years. Most texture synthesis techniques can becategorized into two types: procedural texture synthesis and texture synthesis fromsamples. Procedural texture synthesis renders texture on a curved plane directly bysimulating real-life physical objects such as hair, clouds and wood textures, thus to avoiddistortion brought by texture mapping. But this method demands numerous experimentsto re-adjust system parameters for each new texture, which makes it much lessconvenient. Texture synthesis from samples is to generate a larger area of texture imageby sampling texture in smaller areas according to texture self-similarity. This methodovercomes the drawbacks of procedural texture synthesis, therefore it becomes more andmore popular in textural synthesis research area. Conventional texture synthesis fromsamples often analyze textures statistically, however they can only handle randomtextures. The Markov Random Field (MRF) based pixel-by-pixel synthesizing approachimproves the quality of the synthesized image but is slow. Sample-based texturesynthesis using patch-combination not only increases synthesizing speed, but alsosynthesis high quality images by exploiting neighbor locality of textures. Thismethodology enlarges the realm of textures that can be worked on, and proves to be thedirection of the future.In the texture synthesis research area, the synthesis speed and synthesis quality aretwo criteria to evaluate a successfully texture synthesis algorithm. From the early usestatistical method, to sample-based texture synthesis algorithm in the recent years, thesealgorithms were pursuing quicker synthesis speed and higher synthesis quality. Thispaper analyses from these two aspects, presents separately fast texture synthesisalgorithm using Particle Swarm Optimization (PSO) and Directional Empirical ModeDecomposition (DEMD)-based texture synthesis algorithm. This paper also researchesseveral important applications of texture synthesis.Fast texture synthesis algorithm using PSO is an effective texture synthesisalgorithm. The PSO algorithm is a classical algorithm in swarm intelligencealgorithms which is introduced to the texture synthesis domain. It was discoveredthrough simulation of a simplified social model. The basic algorithm of PSO involvedforming a set of particles over the search space, each with an individual, initially random,location and velocity vector. The particles travel over the search space, remembering thebest fit location experienced. During each iteration, each particle adjusts its velocityvector based on its momentum and the influence of its best location and the best locationof its neighbors. Then it computes a new point to examine. Each Paritcle tends to a local,non-optimal extrema. However, by each particle considering both its own memory andthat of its neighbors, the entire swarm tends to converge on the global extrema.The PSO algorithm is applied to improve the process of searching and matching inthe texture synthesis by patch-based sampling. Although the patch-based real timetexture synthesis algorithm is an accelerating method, it needs three accelerationalgorithms thus very complex to implement. This paper uses the PSO algorithm hasfewer parameters, quicker astringency, and is easier to realize. It speeds up the processof synthesis without influencing the quality of the image. It synthesizes high qualitytextures with a mid-level PC in real-time, and also reduces the algorithm complex toimplement. How the number of the particles and the iterations influence the speed of thesynthesis is analyzed in detail.This paper proposes another novel texture synthesis algorithm: DEMD-basedtexture synthesis, which is aiming at achieving good texture synthesis quality EmpiricalMode Decomposition (EMD) algorithm was proposed by Huang et al in 1998 and it is anovel data analysis method. Its core concept is that any complex data set can bedecomposed into a finite set of Intrinsic Mode Functions (IMF) and a residual variable.Because of EMD's excellent local self-adaptability, it was immediately applied intoone-dimensional non-linear, non-stable data analysis such as ocean tide analysis,earthquake analysis. In recent years, Bidimensional Empirical Mode Decomposition(BEMD) proposed by Nunes et al has been used in multi-measurement analysis and hasreceived positive feedbacks. However, BEMD method lacks considerations in twoperspectives: one is the usage of an image's inherent direction in bi-dimension EMD, theother is extracting and exploiting image signature from bi-dimension EMD. To addressthe above two aspects, Liu et al proposed directional EMD (DEMD), which takes imagedirection into account when decomposing the framework. DEMD also extracts threefeature values from every pixel to do image processing. This algorithm is also used intexture classification and segmentation, and experiments results demonstrated itseffectiveness in texture image processing.This paper proposes an algorithm based on DEMD, which uses DEMD to firstcalculate the inherent direction of the sample texture, and then perform DEMDdecomposition to the sample texture in this direction, thus to obtain a set of IMF imagesand a residual image. By improving the similarity matching procedure for local textures,which is often used in patch-based texture synthesis approach, this algorithm can docorrelation search on feature values for each IMF image generated by DEMD. It also canincrease search information, to seek more matched texture patches. All levels ofsynthesized images will be used as the synthesized image's IMF image and residualimage, and by applying DEMD property to compute the synthesized results for thesample texture in its inherent direction. In the end, the final synthesized texture imagecan be obtained using inverse transformation in the calculated direction from DEMD onthe result image. Compared with other texture synthesis algorithm, our approach notonly maintains the features of the sample texture, but also synthesis higher qualitytexture synthesis through further details in multi-measurement analysis. It is also simple,efficient and can be applied to a wide range of applications.Textures taken from the nature world are often found inadequate to meet therequirements and we need to edit the textures to generate different textures for people'sneed. How to generate new texture from existing texture samples is of great significance.In this paper, we propose several new algorithms of texture synthesis in the imageprocessing area. We present a new texture transfer approach using patch-based texturesynthesis method. Most contemporary texture transfer methods are derived frompixel-based texture synthesis while this method is based on DEMD-based texturesynthesis algorithm. It allows the user to control the transfer process by adjusting theparameters according to the requirements and the features of the textures. This methodimproves transfer speed while keeping transfer quality hence good practicability. In thispaper, we presents a fast image analogy algorithm, which employs patch-based texturesynthesis approaches, as well as speeds up texture synthesis process by exploiting PSOduring matching search process. This algorithm also accomplishes orientation control ofmatching patches in structure and details through appropriate parameter setting. Once acertain sample image with artistic style is presented, the algorithm can accomplish imageanalogies by transferring the style to the target. Our algorithm extracts properties forminput sample using patch-based texture synthesis method, which differs from thepixel-based texture synthesis method. It uses PSO method to search among input texturesamples. Also when searching texture property patches in sample image, instead of ausing traversal search, an optimal matching is found by exploiting random particles.Through these two aspects changes, analogy is improved obviously. In common imageanalogy methods several input sample images are needed, while in our method only theinput sample image is needed. We propose two constrained texture synthesis algorithms.The first one is the constrained single-sample texture synthesis. By extending the fasttexture synthesis algorithm using PSO and applying it to constrained single-sampletexture synthesis, we repaired the blemish. Through the usage of different types oftexture segment to find matching and the texture patch which can change in size, werepaired the boundary of the blemish, making no obvious sense of fringe in the boundaryof the repaired picture. The second algorithm is the constrained two-sample texturesynthesis. We take three images as input: a target image and two texture samples. Thenwe synthesize texture on different areas from different sample textures. Basically wefirst synthesise two temporary textures of suitable size from the sample textures usingthe based-DEMD texture synthesis. Then we render one of the temporary textures to aspecific area according to the user's need. We can design various new graphs accordingto the user's demand;we can also edit the existing photo to attain the seamless mosaicseffect between textures. The two constrained texture synthesis algorithms are fast andare simple to implement.Panoramic image is a method of making use of realistic images to get a full view ofpanoramic space. Users can use ordinary cameras to take a serial of images surroundinga scene. When there is overlap in these images, the system can then automatically createa 360°panoramic image. This can provide users with the ability to observe the virtualenvironment, walkthrough in the virtual environment, and perceive the environmentfrom different viewpoints and directions voluntarily. The panorama can be applied tovarious kinds of fields. For example, in virtual reality, it can be used to replacecomplicated 3D scene modeling and rendering. It is also applied in the field of videocompression, video transmission and medical science, etc. Although the panorama issimple to construct and can represent static scenes naturally, it is unable to showdynamic scenes. Such monotone scenes make the panorama a very noticeable artifact. Inorder to overcome this disadvantage, we hope to add the dynamic scene into thepanorama. An effective method to represent dynamic scenes is therefore needed.Continuous video is a good choice, but the limitation is that it is a very specificembodiment during a very specific period of time. So we need to repeat the video togenerate a "timeless" one. However, this method can lead to visual discontinuitybetween the last frame and the first frame, and bad randomicity may also generate therepeated feeling. Video texture to solve this problem has been proposed. But it cannotprovide panorama or large view.We present a novel method to create dynamic panoramas. We add video texture tothe panorama to implement dynamic panorama. The dynamic panorama keeps theadvantage of providing both full view of scene in static panoramas and the dynamics ofthe scene, which remarkably strengthens the reality of walkthrough. For the staticpanorama, introduces the two-dimensional texture match search thought to the imagemosaic area, unifies multiresolution technology and pattern recognition proposed a newfast image mosaics algorithm PSO-based multi-resolution mosaics algorithms.Compared with other image mosaics algorithms, our approach is straightforward andeasy to implement. We first use PSO to find a certain area which contains sufficientobjective characters, then we use pattern recognition method to search the matchingpatch in another image and adjust image;at last, the mosaic image is created by amulti-resolution method. Experimental results show that this algorithm is able toseamlessly stitch two overlapping images automatically. We develop a novel panoramabrowser, which implements functions such as adding, matching and playing videotextures in addition to its original functions.In conclusion, the achievement of our research results enriches the approaches tothe texture synthesis technology and its applications, and has a certain theoretical andpractical importance. This paper provides useful methods and approaches for theresearch and the development of texture synthesis technology and its applications.