Research And Implementation Of Key Technologies For Image-based Book Spine Detection And Recognition

This paper is aimed at convenience self-help bookcases.It adopts image-based spine detection and recognition method,manages and verifies the operations such as borrowing and returning books,and can reduce the workload of manual management of bookcases and improve work efficiency.This paper combines the functional requirements of self-service bookcases,researches and implements the spine detection and recognition methods suitable for different application scenarios.The main work is as follows:(1)For the scenes in the bookcase grid,a book spine detection method combining line features and text features is proposed to realize the positioning and segmentation of each spine in the bookshelf image.In order to determine the position and orientation of the spine,an approximate line detection algorithm based on chain code tracking is proposed to perform coarse positioning of the spine region.In order to solve the problem of missed detection and false detection of coarse positioning,the CTPN text detection algorithm is used to detect the text column and accurately locate the spine.This method effectively improves the problems of missed detection,false detection,adhesion,etc.in previous research methods.(2)For the multi-scale,multi-position spine detection in natural scenes,the spine segmentation model is trained based on the Mask R-CNN segmentation network,which has better flexibility and versatility for the spine detection in the changing scene.By designing the training set and preprocessing method suitable for the large target spine,the model has improved the spine detection and segmentation effect of medium and large target spines.(3)For application scenarios of automatic library construction,the book recognition method based on character recognition is designed and implemented.For the characteristics of the text color of the spine text,the character segmentation method based on SWT algorithm and projection analysis is designed and implemented.For the rich characteristics of the text of the spine,the Chinese character recognition model based on convolutional neural network was designed and trained.The matching and correction of the text of the spine and the actual book information based on the shortest edit distance algorithm proved that the method has better recognition effect on Chinese character spine than previous works.(4)For application scenarios such as book matching retrieval,borrowing and returning verification,etc.A multi-object template matching method is designed and implemented based on SIFT feature and RANSAC feature matching algorithm.Experiments show that the method can match the target spine image more accurately and conveniently than previous work,so that to effectively verify user compliance with bookcase books.