Research And Implementation Of Automatic Network Texture Acquisition Algorithm Based On Synthesizable Metrics

With the wide application of multimedia information and technology, and the innovation of intelligent era, the network environment provides enough natural textures for people. How to automatically extract the perfect texture examples from the network images, improving the implemented effect of subsequent texture synthesis from the source, it is a new challenge to researchers.Currently, the study of texture synthesis has become a very active area, but researchers are mostly concerned on learning the algorithms of texture synthesis to achieve higher quality and better efficiency of the synthesis. In this paper, we don’t improve traditional texture synthesis algorithms, instead of, by analyzing the texture features, and innovatively summarize three kinds of characteristics based on the synthesizability metric, namely textureness, homogeneity, and repeatability. In this paper, it is enclosed the acquisition algorithms of synthesizable texture examples in-depth study, mainly to complete three aspects of the research work as following:First, we get a network of natural images by web crawler, and use Poisson disk sampling fixed position on a given image globally, which is possible to quickly determine an area as texture. It automatically acquire texture examples based on different sampling dots constantly update iterative. In this process, we have understood the principles of Poisson disk sampling algorithm, how to generate new sampling points, and demonstrate the randomness, uniformity of Poisson disk sampling dots comparing with random sampling.Then, we propose a theoretical framework for the synthesizability measured algorithms of obtaining texture samples, design three characteristics: textureness, homogeneity, and repeatability by analysis of textures. This article describes the principle of the three feature algorithms in detail. Textureness, is mainly by extracting image Gist feature as a parameter, using support vector machine(SVM) to classify texture image and scene image as a trained model, then different samples are carried out in the training model, and ultimately returns the predicted value; Homogeneity, is mainly used K-means clustering algorithm and KNN classification algorithm to process the texture pixel block in the clustering, and calculate the similarity of two equal-sized random cluster areas by the homogeneity algorithm in our approach; Repeatability, mainly use the matching algorithm with normalized cross correlation blended in convolution, to match texture examples self-relativity, calculate random a piece of matching area by the repeatability algorithm in our approach. These three characteristic algorithms for different texture examples corresponding characteristic score, the level of texture scores indicates the degree of the synthesizability.Finally, by the researchful project for obtaining the network texture automatically, we set up a system about three characteristic algorithms of the synthesizability metric of different texture examples in our paper. The system acquires samples can be used for texture synthesis and the corresponding characteristic scores when input natural images, then sorting relatively the high scores of texture examples, the local optimization of post-processing surrounding examples is stored in texture library. Ultimately, we deal with experimental results both in quantitative and qualitative analysis, which testifies the effectiveness and feasibility of three texture characteristic algorithms in our approach.