图书馆在架图书书脊与标签实例分割方法研究  

Research on Instance Segmentation Method between Spine and Label of Library Books on Shelf

作　　者：尤永鹏 董峦[1] 尹书林 李佳航 艾里亚尔·阿不都克里木[2] YOU Yong-peng;DONG Luan;YIN Shu-lin;LI Jia-hang;Ailiyaer Abudukelimu(College of Computer and Information Engineering,Xinjiang Agricultural University,Urumqi 830052,China;Library of Xinjiang Agricultural University,Urumqi 830052,China)

机构地区：[1]新疆农业大学计算机与信息工程学院,乌鲁木齐830052 [2]新疆农业大学图书馆,乌鲁木齐830052

出　　处：《新疆农业大学学报》2023年第2期139-145,共7页Journal of Xinjiang Agricultural University

摘　　要：准确获取图书信息是智能化管理图书的关键,为实现在架图书书脊和书脊底部标签的精确分割,本研究在SparseInst基础上提出增强编码器实例分割模型EE-SparseInst。该模型使用特征选择对齐模块将特征图对齐并融合,以减少边界信息损失,使用边界解析模块增强相邻书脊边缘处的语义表示,提高掩码的分割质量。建立在架图书图像数据集,数据集包含2253张图像和标注信息。结果表明,EE-SparseInst平均精确率达到80.81%,相较于CenterMask、SOLOv2、PolarMask和SparseInst 4种主流锚框自由式实例分割方法分别提高了7.57%、20.07%、10.50%、0.74%。Obtaining the information of books accurately is the key of intelligent management of books,so in order to achieve the accurate division of the spine and the bottom of the spine label of the shelf book,this paper proposed the Enhanced Encoder instance segmentation model EE-SparseInst based on SparseInst.The feature selection alignment module was used to align and fuse feature maps to reduce boundary information loss and the boundary resolution module was applied to enhance the semantic representation at the adjacent spine edges,thus improving the segmentation quality of the mask.In addition,an on-shelf book image dataset was constructed,which contained 2255 images and annotation information.The experimental results showed that the average precision of EE-SparseInst reached 80.81%,which improved 7.57,20.07,10.50,and 0.74 percentage points compared with the four mainstream anchor frame free-form instance segmentation methods:CenterMask,SOLOv2,PolarMask,and SparseInst,respectively.

关 键 词：实例分割 图书分割 特征对齐 边界增强 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]