Research on Library Book Spine Segmentation and Recognition Based on Deep Learning

With the rapid development of artificial intelligence technology, library management has become increasingly intelligent. Addressing the deficiencies of existing spine segmentation and recognition methods in terms of accuracy and utilization of fine-grained visual information, this paper proposes an improved algorithm based on the Mask Region-based Convolutional Neural Network (Mask R-CNN) model. Specifically, the SE (Squeeze-and-Excitation) channel attention mechanism is introduced into the C4 and C5 stages of the ResNet-50 backbone network, and a phased training strategy is adopted. Experimental results demonstrate that the proposed Mask R-CNN (SE-C4C5) model achieves optimal performance in both spine detection and instance segmentation tasks, with significant improvements especially at higher IoU (Intersection over Union) thresholds. Meanwhile, an integrated OCR interface enables fast and accurate recognition of the text on segmented spines. This method effectively enhances the accuracy of spine recognition and provides a feasible technical solution for automated inventory and intelligent management of libraries.