Research And Implementation Of Automatic Network Texture Exemplar Extraction Algorithm

With the rapid development of the Internet,people could obtain lots of texture images through various crawler software and search engine.Ideal texture exemplar could inspire artists in daily creation,help designers to produce more realistic scene,assist doctors in medical diagnosis,and provide the evidence of species classification,etc.However,conventional method on texture exemplar extraction is still a labor intensive task.After collecting a large number of texture images,it requires frequent manual operations,such as screening,cropping and editing.With the advent of the era of artificial intelligence,it's becoming increasingly urgent as automatic extracting replace manual screening.Conventional texture analysis mainly focuses on the regional feature analysis and signal processing method.After decades of development,these analytical methods have achieved desired results in the fields of image synthesis,segmentation and classification.With the advent of the era of artificial intelligence and big data,the convolution neural network has been widely used in image feature analysis.In this paper,we respectively employed conventional method and the convolutional neural network to automatic texture exemplar extraction.According to this research,people could emancipate themselves from annoying and burdensome screening.The proposed method could concluded as the following three sections:Firstly,for texture images obtained from the Internet,we could uniformly crop dozens of unevaluated exemplars by region proposal methods such as Poisson sampling.Each exemplar was been evaluated by traditional texture analysis methods,which we proposed as image homogeneity,i.e.regional feature clustering analysis.Region similarity was been assessed by counting the number of cluster categories in the equal size area.With small area of non-textured or mixed texture,defective texture could obtained high scores due to its higher global homogeneity.We redefined the desired texture scores by texture homogeneity and defect filter algorithm.In this way,the ideal texture exemplar was raised to the top.Secondly,although conventional method could automatically extract texture exemplar,its single feature representation could not generate more extensive texture categories.Meanwhile,applying Deep Learning technology into automatic texture exemplar extraction is a fresh topic.The region proposal method of automatic texture exemplar extraction have much common with object recognition.But the ideal texture exemplar doesn't have definite boundary.It's difficult to locate the position of ideal exemplar by bounding box regression.In this paper,we ported selective search algorithm from objective recognition to ideal exemplar extraction,which turn out to have more recommendatory region proposal.Different from the existing attribute-based texture library,we categorize texture by its structure,i.e.regularity and sparsity.Then,we could got more comprehensive and more complete texture represent method.Finally,according to the methods above,we constructed a high operability automatic ideal exemplar extraction system.Once input a source images,we could obtained dozen of ideal texture exemplar by selecting different algorithm.Each exemplar was sorted by its final scores.For more convincing results,we employed our method on thousands of source images.Visual results indicated that our proposed method could accurately extract texture exemplars from arbitrary source images.