Research On Image Detection And Retrieval Model Of Silk Cultural Relics Under The Framework Of Deep Learning

Cultural relics are the precious history and culture heritage for each country.And Chinese culture has a long history and countless precious culture relics have been left over for thousands of years.These culture relics can not only motivate current people to explore ancient life,but also greatly enhance the pride of the Chinese nation.Therefore,it is particularly important to identify and protect culture heritage nowadays.Recently,with the promotion of nation policy,i.e.one belt,one road,the whole country becomes paying more attention on silk cultural relics.Silk is the motive force of the silk road.It has been spread all over the world through the Silk Road and contributes a lot to human civilization.However,due to the wide distribution of silk heritage and complex temporal and spatial relations,there are serious deficiencies in the investigation and scientific cognition of the origin region of silk,the propagation route and mode of the silk road.This will cause some obstacles on the discussion on the spatial-temporal law of silk propagation and communication.On the other hand,the image-based cultural relic identification technology has made some progress due to the development of computer vision technology.However,due to the lack of a unified cultural relic image acquisition and search platform,the collected data cannot be better shared among different research teams.This will result in repeatedly labelling and cause a waste of human laboring,so that the normal study on the propagation law of silk cultural relics will be blocked.To handle the above problems,we in this thesis firstly collect special datasets for the study on the silk cultural relic images.We then propose a deep learning model for detecting,tracking and retrieving silk cultural relic pattern.Finally,we design and develop a system for integration of the above tasks.In practice,this system is firstly to detect and track the excavated silk cultural relic patterns through the camera.It is then to retrieve the detected patterns in the database.through which a data collection and searching platform for silk cultural relic pattern can be provided for relevant researchers.In general,the main work in this thesis is as follows:1.We have finished the collection of silk cultural relic images.In detail,we in this work collect various categories of silk cultural relics pattern images from specific field.We then select some categories of silk cultural relic pattern that can best represent a specific period or a certain decorative style.Based on the selected data,we finally formulate the silk cultural relic pattern detection dataset and silk cultural relic pattern retrieval dataset,respectively,that can be used in the following experiment.2.We have developed an improved YOLOv5 network for detecting silk cultural relic pattern.This model has two improvements compared with original YOLOv5 network: To deal with the problem of detecting corrupted and small-target pattern,the proposed model have adopted several strategies,such as expanding the target detection layer,inducing the adaptive activation function and Coordinate Attention mechanism,for enhancing the detection ability for corrupted and smalltarget patterns;To deal with problem of detecting dense patterns,we have replace the original NMS with Diou Loss based NMS,,which can effectively remove the redundant detection boxes generated in the detection procedure.Based on his above improvements,the overall detection Map based on the silk cultural relic pattern detection dataset can increased from 0.958 to 0.966 compared with original YOLOv5.The presence of redundant detection boxes can also be greatly reduced in the on-line detection procedure.3.We have developed a model for tracking the silk cultural relic pattern by combining improved YOLOv5 network with Deep SORT network.This model is to tracking detected silk cultural relic pattern based on the Deep SORT model and to calculate the pattern number in the video..We also compare the tracking performance of the proposed model with other state-of-theart methods such as DETR+Deep SORT methods.4.Under the framework of deep learning,we have realized the silk cultural relics image retrieval network model based on dense residual network and N-Pair-Loss.Compared with the traditional retrieval model,the proposed model can achieve higher retrieval performance with small samples,which indicates that the proposed model is suitable for handling small-sample retrieval problem of silk cultural relic pattern images.5.We have design and develop a set of detection,tracking and retrieval system based on the above network for silk cultural relic pattern.The system is mainly divided into two parts: image detection and tracking part and retrieval part.The detection and tracking part are to collect silk cultural relic images via an external camera,and then to detect and track silk cultural relic pattern from the collected images;the retrieval part is to crop the detected silk cultural relic patterns and to retrieve them in the database.The detection and retrieval system can provide a searching platform for relevant researchers on the study of silk cultural relic pattern image.