Abstract: Due to the arriving of the Internet plus era, the number of online images have increased dramatically, and content-based image retrieval has attracted a lot of attention. Traditional retrieval methods are inefficient because of the low expression ability, which is not conducive to large-scale image retrieval. So, a new image retrieval algorithm based on convolutional neural network is proposed. A new end-to-end convolutional neural network framework is designed which learns probability based semantic-level similarity and feature-level similarity simultaneously. The principal component analysis method is introduced to reduce the dimensions of deep features and reduce the loss of information. The distance function is used to calculate the distance between the target image and the database image to realize the retrieval. Experimental results of ImageNet-1000 and Oxford 5K datasets indicate that the expression ability of visual feature is effectively improved and the image retrieval performance is substantially boosted compared with the contrast methods．

Key words: image retrieval, convolutional neural network, principal component analysis, deep features

摘要： 由于互联网+时代的到来，在线图像的数量急剧增加，基于内容的图像检索引起了很多关注。传统的检索方法由于图像表达能力不强，使得检索效率低下，不利于大规模图像检索。因此，提出一种新的基于卷积神经网络的图像检索算法。设计一种新型的端到端的卷积神经网络结构，同时学习基于概率的语义信息相似性和图像特征相似性；引入主成分分析方法，对深层特征进行降维的同时降低信息的损失；通过距离函数计算目标图像与数据库图像的距离，实现检索。在Image Net-1000和Oxford 5K数据集上的实验结果表明，该方法能够有效地增强图像特征的表达能力，提高检索性能，优于对比方法。

关键词: 图像检索, 卷积神经网络, 主成分分析, 深层特征