Application Research Of Deep Learning In Embroidery Image Retrieval

Embroidery images,as one of the unique intangible cultural heritages in Qinghai Province,have the most distinctive national characteristics and regional characteristics among many folk arts in Qinghai.Embroidery images have important research value as the carrier of embroidery art content information.In recent years,image retrieval has become the most important research problem and hotspot in the image-related industries.At the same time,with the development of the Internet,deep learning has achieved remarkable results in the research of image information feature extraction,classification,recognition,retrieval,etc..In order to enrich the means and methods for the digital protection of Qinghai embroidery art,this paper designs and implements an embroidery retrieval system based on the theory of deep learning.On the one hand,it makes up for the gap in the research of embroidery image retrieval,so that users can quickly retrieve what they need The embroidery image information on the other hand also played a positive role in the digital protection,inheritance,development and utilization of Qinghai embroidery art.The main work of this paper is:(1)This thesis combines deep learning with embroidery image retrieval technology,puts forward the application model of deep learning in embroidery image retrieval system,and designs a reliable embroidery image retrieval system.This paper solves the problems of lack of learning ability and slow retrieval speed in traditional text-based retrieval methods and traditional content-based retrieval methods.(2)Starting from the function and performance of the image retrieval system,this paper uses embroidery images as data sets to complete the analysis and design of the embroidery image retrieval system,including the construction of the system platform and the realization of the system function modules.The embroidery image retrieval system uses the Caffe framework and AlexNet network model based on deep learning.In the process of image retrieval,the embroidery image data set is first preprocessed,and the processed data set is model trained through the AlexNet network model.Use the trained model to extract the features of the embroidery image,measure the similarity distance between the images through the Euclidean distance,and finally sort according to the similarity size,and select the top K image output results and image information with the smallest similarity distance.(3)In the experiment of this paper,while performing model training,a fine-tune strategy is proposed for the ALexNet network model.While improving the existing model training,by updating some network-level weights,Retrain the last layer of the network structure,and repeatedly train the data set to achieve the desired desired effect.In order to compare the effects,this paper uses the AlexNet network model and the finetuned network model to extract the features of the embroidery images.By measuring the similarity between the images,the Top K results sorted according to the similarity are finally returned,that is,the top K images And its image information.The experimental results show that the network model after fine-tuning strategy is more accurate and efficient in the embroidery image retrieval system.