Fast-robust book information extraction system for automated intelligence library

At present, in the large-scale book management scene, book sorting, daily maintenance and book retrieval are very common, but the book information is complicated and the efficiency of relying on manual management is extremely poor. Although there have been many self-service book systems based on optics or vision, they are mostly based on traditional computer vision algorithms such as boundary extraction. Due to the fact that there are more artificial experience thresholds, some shortcomings such as low detection accuracy, poor robustness, and inability to systematically deploy on a large scale, which lack of insufficient intelligence. Therefore, we proposed a book information extraction algorithm based on object detection and optical character recognition (OCR) that is suitable for multiple book information recognition, multiple book image angles and multiple book postures. It can be applied to scenes such as book sorting, bookshelf management and book retrieval. The system we designed includes the classification of book covers and back covers, the classification of books upright and inverted, the detection of book pages side and spine side, the recognition of book pricing. In terms of accuracy, the classification accuracy of the front cover and the back cover is 99.9%, the upright classification accuracy of book front covers is 98.8%, the back cover reaches 99.9%, the accuracy of book price recognition get 94.5%, and the book spine/page side detection mAP reaches 99.6%; in terms of detection speed, Yolov5 detection model was improved and the statistical-based pre-pruning strategy was adopted, support by our algorithm the system reaches 2.09 FPS in book price recognition, which improves the detection speed to meet actual needs.