Research On Key Technologies Of Visual Book Inventory System

With the increasing number of books and readers in the library,intelligent management becomes more and more important.Especially in the book inventory task,more and more technologies are applied to improve the efficiency of book circulation.In recent years,thanks to the development of deep learning and computer vision technology,the visual-based book counting system has attracted the attention of many researchers because of its low cost.However,most of the existing visual schemes assume that the spine is placed orderly,and the damage of the spine is not considered,resulting in unsatisfactory results in practical application.This paper studies the key technologies in the visual book inventory system,and further improves the performance of the inventory system.The specific research contents are as follows:First of all,the damaged or poorly placed spine will seriously affect the subsequent segmentation and recognition,which will lead to false positives.Therefore,this paper first introduces a quality discriminator in the front end to filter the images with damaged or poorly placed spine,which can effectively reduce the false alarm rate.In view of the phenomenon that existing segmentation schemes cannot effectively segment dense slanted book ridges,this paper introduces angle bounding boxes and mask evaluators based on the original Mask RCNN model to obtain an improved model Library Net.Compared with existing segmentation schemes,this model can significantly improve the segmentation performance of dense slanted book ridges.Then,aiming at the phenomenon that the existing schemes cannot effectively detect the spine characters,this paper proposes a detection model Spine OCR by improving the EAST model.Experiments show that its text detection performance is better than the existing schemes.In view of the phenomenon that the spine character data set is short,which leads to the low accuracy of character recognition,this paper uses artificial synthesis and crawling the book cover of DANGDANG to achieve data amplification,and uses the augmented data to train the CRNN model,which has greatly improved the accuracy of spine character recognition.For highly similar series books,the feature retrieval accuracy of existing schemes is generally low.This paper refers to the idea of single sample learning of commercial face recognition systems,and proposes feature extraction models Spine Siamese and Spine Triplet based on the Siamese and Triplet architecture.The experimental results show that they can effectively learn the fine-grained features of book ridges and improve the retrieval accuracy of series books.Finally,in order to verify the effectiveness of the research work,this paper developed a simple inventory system by integrating the existing programs in the market,including wheeled robot ontology,background system and librarian client software.This simple inventory system integrates all the research results of this paper and has been tested in a small scale in the library of Guangdong Normal University of Technology.The test results show that the research results of this paper have high practical value.