Research And Implementation Of Traditional Patterns Image Retrieval Method Based On Target Location

Traditional pattern is carrying a large number of cultural connotation of the information,is the symbol of traditional culture,in the era of digital intelligent,cultural calculation plays an important role in mining and inheriting the connotation of traditional culture.Along with the traditional culture research in the process of the growth of the accumulation of image data,and the diversity of traditional pattern carrier result in different effects on different carrier,Image retrieval accuracy is also affected by the carrier,and how to quickly and accurately find the original image of the pattern extracted from the image,to understand the cultural context and scene of the traditional pattern has become a difficult problem,how to apply computing and related technology to the cultural field,to achieve the image retrieval of the traditional pattern has also become a problem to be solved.Based on traditional pattern as the research object,this paper aiming to realize the subgraph of traditional pattern with the sketch retrieval,study of image retrieval technology,based on the Transformer build sub-image retrieval and hand-painted sketch retrieval algorithm.And based on the proposed algorithm,traditional pattern sub-image and sketch image retrieval system is built.In this paper,the main research contents and innovation points are as follows:(1)Construct the data set of traditional pattern sub-image and sketch image.In this paper,books,PatternNet resource library and the Internet are the sources of the data set.By choosing the traditional Chinese pattern books,using scanning device scan chosen book image,and collating existing data sets in the laboratory,the collected set of data is preprocessed,segmented to extract the sub-image,and finish the corresponding sketch,build in the traditional pattern sub-image and sketch data set.There were 3074 images in traditional pattern sub-image data sets and 1192 iamges in sketch data sets.(2)Build a sub-image retrieval algorithm named SIRT based on Transformer.This paper studies the traditional pattern sub-image retrieval technology and proposes a sub-image retrieval model based on selfattention mechanism.Feature maps of different layers of CNN are extracted and fused by feature pyramid to obtain fusion feature maps with both good spatial information and obvious semantic information.Database images are extracted global feature maps and local feature maps with the prediction box of Transformer,then feature aggregation algorithm is used for aggregation.Since the user input is extracted from the original image,the background in the image has been removed,and only specific areas in the sub-image are helpful to construct differentiated global features,so Vision Transformer is used to improve the R-MAC aggregation algorithm in the sub-image feature extraction stage,and the experiment proves that the module can effectively suppress invalid features.Greatly improve the retrieval accuracy..(3)Build a sketch-based image retrieval algorithm named Triplet-VIT based on Triplet.This paper proposes a sketch retrieval method based on Transformer.Freehand sketch is an abstract representation of natural image,with only a small amount of information such as contour shape,and there is a huge difference between it and natural image.In order to solve the cross-domain problem between freehand sketch and natural image,the edge contour of natural image is extracted first,and then the metric learning method of semi-weight sharing is used to narrow the differences between the two domains.Nowadays,Transformer has made outstanding achievements in many natural language processing tasks.Recently,Transformer has also been successfully applied in computer Vision.This paper combines Transformer with CNN for sketch retrieval,using Vision Transformer to encode the feature images extracted from CNN to generate image descriptors,after training.(4)Design and implement the image retrieval system.Based on the proposed retrieval model of traditional pattern sub-image and hand-drawn sketch,the traditional pattern image retrieval system was built to realize the retrieval of the traditional pattern sub-image and sketch.This paper takes traditional patterns as the research object and aims to achieve efficient and accurate traditional pattern sub-image retrieval and sketch image retrieval.A set of traditional pattern sub-image and sketch image data set are constructed.A sub-image retrieval algorithm based on Transformer and a sketch retrieval algorithm based on Triplet are proposed.Finally,the proposed algorithm is integrated with the traditional pattern image retrieval system.