基于改进Mask R-CNN的在架图书书脊图像实例分割方法  

Improved Mask R-CNN based instance segmentation method for spine image of books on shelves

作　　者：曾文雯[1] 杨阳 钟小品[2] Zeng Wenwen;Yang Yang;Zhong Xiaopin(Library,Shenzhen University,Shenzhen Guangdong 518060,China;College of Mechatronics&Control Engineering,Shenzhen University,Shenzhen Guangdong 518060,China)

机构地区：[1]深圳大学图书馆,广东深圳518060 [2]深圳大学机电与控制工程学院,广东深圳518060

出　　处：《计算机应用研究》2021年第11期3456-3459,3505,共5页Application Research of Computers

基　　金：深圳市科技计划项目(JCYJ20180305123922293);深圳大学—台北科技大学学术合作专题研究项目(2019009)。

摘　　要：运用人工智能技术将是构建下一代智慧图书馆的关键,为了实现图书的定位和识别,提出一种基于改进Mask R-CNN的在架图书书脊图像实例分割方法。考虑到图书密集排列、具有一定的旋转性、副本纹理极相似等难点,改进锚框为旋转矩形框,提出旋转区域建议网络取代区域建议网络;提出旋转特征提取方法可减少池化误差且有效提取目标特征,结合掩膜的旋转对齐以提升预测掩膜的准确性。建立了一个包含1849张在架图书书脊图像的标注数据集,提出方法的测试结果大幅度优于其他重要的实例分割算法,证实了在网络中使用旋转特征对于具有一定朝向的、密集的目标分割难题很有效。The application of artificial intelligence technology will be the key to the establishment of the next generation smart library.In order to realize the location and recognition of books,this paper proposed an instance segmentation method for spine image of books on shelves based on improved Mask R-CNN.Considering the difficulties such as the dense arrangement of books,certain rotation and similar texture of book copies,it improved the anchor as a rotating rectangle box,and also developed a rotating region proposal network to replace the region proposal network.Then it used a rotation feature extraction method to reduce the pooling error,and effectively extracted the features of target with rotation,combined with a rotation alignment of the mask to improve the accuracy of the prediction mask.This study built a labelled dataset containing 1849 spine images of on-shelf books.The test results of the proposed method are much better than those of other key instance segmentation algorithms.It is proves that the use of rotating features in the network is very effective for instance segmentation of dense targets with orientation.

关 键 词：智慧图书馆 图像分割 Mask R-CNN 旋转特征提取 

分 类 号：TN911.73[电子电信—通信与信息系统]