Research On Automatic Recognition Technology Of On-shelf Book Based On Vision

With the development of information society,the public demand for library resource is improving explosively.Traditional document center model based on resource is turning to the reader center model leaded by demand.Affected by the intelligent technology,software and hardware getting more and more mature,the development of library management system is changing toward the direction of intelligence.Therefore,this study in the library automatic identification technology has important theoretical significance and practical value in production.Today,machine vision applications covering industry,agriculture,transportation,military,aerospace and other fields,the analysis judgment by machines instead of human eyes get rapid development.So adopted the machine vision technology in this paper,with accurate segmentation for spine and recognition for call number as the foundation to identify books to implement in books automatic identification as the main research target,we propose an automatic identification technology for the wrong order books based on machine vision.As to segmentation of book spine,we believe the spine boundary show as a straight line in the image as prior conditions,using the line detection strategies to segmentation of books.In addition,the existing of line detection algorithm based on Hough transform and LSD caused a series of problems,such as mistakenly identified,time efficiency and limitations of these algorithms,this paper proposes a fast line detection algorithm based on edge analysis and direction constraints.This algorithm combines the advantages of the Canny operator and digital line features.First,perform the image edge detection.Then,track the edge image to build the line chain based on digital line feature.Finally,connect two endpoints of extracted line chain to get the complete line that is book spine.At the same time,compared to mainstream line detection algorithm,the experiments show that the proposed algorithm is more efficient and accurate in line detection,and very suit to real-time image processing,machine vision applications,etc.In addition,this article adopts the deep learning technology,which has developing in recent years to design a continuous double convolution network model for call number identification.Compared with the traditional network structure,this model is made of two consecutive 5×5 convolution kernel for call number image feature extraction.Uses unsaturated nonlinear function1)()= max(0,)to simulate neurons,so that solve the problem of gradient disappear and reduces the network training time.Bring in the Dropout model integration strategy to improve the accuracy and stability of the model and effectively avoid the fitting situation.Above all,solved the problem of lower recognition rate for on-shelf books based on color analysis,feature point extraction and the traditional technique of OCR.