Abstract:

Recently, many scholars have done research on the methods of image processing by combining deep neural networks. Compared with the traditional hand-crafted feature, the feature extracted by deep convolutional networks is more accurate and better. Therefore, this paper proposes a deep network architecture combined with convolutional neural networks and Hashing method for large-scale image retrieval. On the one hand, in order to preserve semantic similarity between binary Hash codes as much as possible, this paper introduces the multi-task learning that the similarity information and the classified information are used to train the network. Moreover, according to the information entropy theory, Hash coding is maintained as evenly distributed as possible to increase the amount of information during the training process. On the other hand, this paper proposes a multi-scale fusion pooling method (MSFP), which can fuse the regional information of various scales, and improve the retrieval performance. Meanwhile, the network parameters are obviously reduced by the pooling strategy. Extensive experiments on three large scale datasets SVHN, CIFAR-10 and NUS-WIDE show the improved performance of this method compared with the state-of-the-arts.

Key words: image retrieval, convolutional neural networks, pooling, Hashing

摘要：

近期，许多学者开始结合深度网络进行图像处理，与传统的基于人工抽取的特征相比，由深度卷积网络提取的特征更为准确，效果更好。因此，提出了一种结合卷积神经网络和哈希算法的深度网络架构，用于大规模图像检索。一方面，为了更好地保持哈希编码之间的语义相似性，引入了多任务学习机制，将图像分类信息和图像间的相似度信息同时用于模型的训练，并且根据信息熵理论，训练过程中使哈希编码尽可能地维持均匀分布以增加信息量；另一方面，提出了一种多尺度融合池化方法（multi-scale fusion pooling，MSFP），融合图像中多种尺度的区域信息，提升了检索性能，同时明显地减少了网络参数。在SVHN、CIFAR-10和NUS-WIDE等数据集上的实验结果表明，提出的算法与现有的基准算法相比，检索效果具有明显改善。

关键词: 图像检索, 卷积神经网络, 池化, 哈希编码