无人图书馆的图书摆放具有极大的随意性，导致现有的书脊分割算法的精度和适应性均无法达到实用的要求，同时大数据量图书的匹配算法也存在实时性差且准确度低的问题，针对以上问题，提出了一种将SAM（segment anything model）分割大模型与多深度特征级联匹配算法相结合的书脊分割与匹配算法，在确保书脊准确分割的前提下对书脊进行精准匹配。首先，引入SAM分割大模型对书籍图像进行分割，其模型通过ViT（vision Transformer）对书脊图像进行高效特征提取，并利用自注意力机制捕捉局部和全局信息，生成高质量的多尺度书脊特征图，从而对书籍图像进行有效准确的分割；其次，提出特征分离网络FHNet（feature hemisection network）模型能同时提取书脊的低频特征、高频特征和高低频融合特征，分别关注书脊图像的颜色、纹理、轮廓信息，增强书脊特征的区分性；同时，引入FAISS（facebook AI similarity search）框架，并将颜色分类、FHNet提取的低频特征、高频特征和高低频融合特征与VGG（visual geometry group）特征做级联匹配，提升算法的实时性。实验表明，多种深度特征级联匹配算法在不同系列书脊匹配任务中准确率为97.93%，在相同系列书脊匹配任务中准确率为95.48%，匹配检索时间为0.3 s，其准确率与匹配速度显著优于其他书脊匹配算法，进一步证明该算法的有效性。

The random placement of books in unmanned libraries poses challenges to the accuracy and adaptability of existing spine segmentation algorithms, while large-scale matching algorithms are hindered by poor real-time performance and low accuracy. In order to solve those issues, a book spine segmentation and matching algorithm was proposed, where the SAM(segment anything model) segmentation model was combined with a multi-depth feature cascade matching approach to ensure accurate spine segmentation and precise matching. Firstly, the SAM model was employed for segmenting book images, and a vision Transformer (ViT) was used to efficiently extract features. Local and global information was captured through the self-attention mechanism to enable the generation of high-quality multi-scale spine feature maps for effective segmentation. Secondly, the proposed feature hemisection network (FHNet) was utilized to extract low-frequency, high-frequency, and fused features, which respectively focused on color, texture and contour information to enhance spine distinctiveness. Additionally, the FAISS(facebook AI similarity search) framework was incorporated for cascade matching, where color features, FHNet features and VGG(visual geometry group) features were integrated to improve real-time performance. It has been demonstrated through experiments that the proposed method achieves an accuracy of 97.93% in matching tasks across different series and 95.48% within the same series with a retrieval time of 0.3 seconds. The accuracy and matching speed are significantly improved compared to other algorithms, further validating the effectiveness of this approach.