基于卷积神经网络和哈希编码的图像检索方法

摘要

在图像检索中，传统的基于人工特征的检索方法并不能取得很好的效果。为此提出一种结合卷积神经网络和以前最好水准的哈希编码策略的图像检索方法。鉴于近几年卷积神经网络在大量的计算机视觉任务上的巨大进步，该方法首先使用在ILSVRC数据集上预训练过的VGGNet－D网络模型对实验图像数据集提取卷积特征来得到图像的深层表示，再采用以前最好水准的哈希策略将这些深层表示进行编码，从而得到图像的二进制码，最后再进行快速图像检索。在两个常用的数据集Caltech101和Caltech256上的实验结果表明，本文方法的5个策略相比于以前最好水准的相应的图像检索策略在“精度－召回率”和“平均正确率值－编码位数”两个指标上能获得更优异的性能，证明了本文方法在图像检索上的有效性。%For image retrieval, traditional retrieval methods based on artificial features are not effective enough. Hence, we propose an image retrieval method, which combines a convolutional neural network and previous state⁃of⁃the⁃art hash coding strategies. In view of the great progress that convolutional neural networks have made in a large number of computer vision tasks in recent years, this method first uses the model"VGGNet⁃D" pre⁃trained on the ILSVRC′s dataset to extract the convolutional features from experimental image datasets to get the deep repre⁃sentations of images, then adopts previous state⁃of⁃the⁃art hash coding strategies to encode the deep representations to obtain the binary codes, and, finally, performs a quick image retrieval. The experimental results on the common⁃ly used Caltech101 and Caltech256 datasets show that this method′s five strategies, compared with the previous state⁃of⁃the⁃art image retrieval strategies, can obtain better, indeed excellent, performance in both the"Precision⁃Recall" and"mean Average Precision⁃Number of bits" metrics, proving the effectiveness of the proposed method in image retrieval.