利用特征融合提升深度学习图像检索算法  

The improvement of the deep learning image retrieval algorithm based on feature fusion to

作　　者：廖逍 王兴涛 徐海青 Liao Xiao;Wang Xingtao;Xu Haiqing(State Grid Information and Communication Group Co.,Ltd.,Beijing,100021,China;Anhui Jiyuan Software Co.,Ltd.,Anhui Hefei,230088,China)

机构地区：[1]国网信息通信产业集团有限公司,北京100021 [2]安徽继远软件有限公司,安徽合肥230088

出　　处：《机械设计与制造工程》2023年第1期112-116,共5页Machine Design and Manufacturing Engineering

基　　金：国网信息通信产业集团项目(5268002XXX38)。

摘　　要：基于卷积神经网络提取图像特征的方法被广泛应用到图像检索中,主要研究内容为设计良好的特征提取方式。为了提高图片全局特征评估检索准确率,对基于特征融合的深度学习图像检索算法进行分析。通过对不同特征提取方式进行测试,提出融合不同卷积层进行特征提取的策略,并且对提取效果进行分析。测试结果显示,检索准确率比单层卷积层提取特征的准确率要高;利用注意力机制融合特征发现通道信息的注意力机制可以提高检索准确率,而空间信息的注意力机制会降低原始信息的可区分度和检索准确率。The method of extracting image features based on convolutional neural network is widely used in image retrieval,and the main research content is the well-designed trait-intensive extraction methods.To improve the average accuracy of image global feature evaluation,the deep learning image retrieval algorithm is analyzed.By testing different feature extraction methods,the strategy of integrating different convolution layers is proposed,and the extraction effect is analyzed.The test results show that the average retrieval accuracy is higher than that of the single-layer convolutional layer,increasing by nearly 2 percentage points;the attention mechanism of discovering the channel information can improve the retrieval accuracy,while the fusion of the spatial information will reduce the discriminability of the original information and reduce the retrieval accuracy.

关 键 词：图像检索 卷积神经网络 特征融合 软融合 注意力机制 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]