Algorithm for spine segmentation of books in naturally placed state

Visual based book positioning is one of the important technologies for automatic book inventory, and the spine segmentation algorithm is the key technology to achieve this goal. Current algorithms struggle with increasing the accuracy and adaptability to the arbitrary arrangement of books in libraries. In order to address this challenge, an improved spine segmentation algorithm based on Deeplab v3 plus is proposed to achieve accurate spine segmentation of books at various tilt angles. First a YOLOv5 detection network is combined with an affine transformation to determine the placement status of books and normalize the tilt angle of spine images. And then the module is integrated into the Deeplab v3 plus framework to form a novel segmentation frame. In order to extract dense and long strip-shaped spine features, a Dense RA-ASPP framework with a large receptive domain and multiple scales is used to replace the ASPP framework in Deeplab v3 plus. Additionally, a CBB module is incorporated to enhance feature extraction and reduce information loss. Experimental results show that the proposed algorithm improves the mean Intersection over Union (MIoU) by 1.58 percentage points, reaching 88.33% on the same test spine database. Moreover, it demonstrates significant advantages in segmenting tilted spine of books and the network's effectiveness in addressing segmentation issues in naturally arranged books in real-world libraries.