摘要
包含丰富信息的图像数据是重要的数据来源,然而要在海量图像数据中进行高效的信息检索却面临不少挑战。通过图像丰富的语义特征区分对象间的差异,图像结构化后的特征维度会达到一个较高的维度,将面临检索代价高、过拟合等一系列问题。不同算法在处理不同类型语义特征的图像数据时,算法的准确性和稳定性有较大差异。为了提高非特定类别图像相似度评价及检索的准确性,设计一种基于集成学习的图像相似度评价方法。首先,以文献中的图像为重点研究对象,从多个权威文献库中收集并提取了大量的图像样本构建样本集;然后,通过对样本集的学习获得了多种评价方法的学习权重,并以此为基础构建基于权重策略的集成学习机制,以提升图像相似度评价方法的稳定性和准确性;最后,通过在样本集上进行测试验证了评价方法的有效性。
The informative image data is an important data source. However,there are many challenges for efficient information retrieval in massive image data. Since the distinguishing of the differences between objects by the rich semantic features of the image,the feature dimension of the structured image will reach a higher dimension,which is faced with a series of problems,such as high retrieval cost and overfitting. When different algorithms are used to deal with image data of different types of semantic features,the accuracy and stability of the algorithms are quite different. An image similarity evaluation method based on ensemble learning is designed to improve the similarity evaluation and retrieval accuracy of images in nonspecific category. The images in the literature are taken as the key research object. A large number of image samples are collected and extracted from multiple authoritative literature library to construct a sample set. And then,the learning weights of various evaluation methods are obtained by the learning of the sample set. On this basis,an ensemble learning mechanism based on the weight strategy is constructed to improve the stability and accuracy of the image similarity evaluation method. The validity of the evaluation method has been verified by testing on the sample set.
作者
熊运鸿
袁琪
王骏巍
卢慧琪
马雪琴
张天凡
XIONG Yunhong;YUAN Qi;WANG Junwei;LU Huiqi;MA Xueqin;ZHANG Tianfan(Hubei Engineering University,Xiaogan 432000,China;Hubei Polytechnic Institute,Xiaogan 432000,China)
出处
《现代电子技术》
2023年第5期74-77,共4页
Modern Electronics Technique
基金
湖北省教育厅科研项目(D20212701,B2020385)
大学生创新创业训练计划项目国家级项目(202110528007)。
关键词
图像相似度评价
集成学习
样本集
投票机制
权重学习
图像特征
image similarity evaluation
ensemble learning
sample set
voting mechanism
weight learning
image feature
