深度哈希卷积网络在图像检索中的应用

Application of Image Retrieval Based on Deep Hash Convolutional Neural Network

随着网络图像数据的迅猛增长,基于传统的人工设计视觉特征的图像检索方法的效果已无法满足要求。为此,提出了一种基于深度卷积神经网络的监督哈希编码方法。该方法是基于分类和量化误差的深度哈希,同时学习图像、哈希码和分类器的特征表示,然后利用标记的监督信息和深层架构学习卷积网络差异性特征表示,并直接用于生成哈希码和预测图像标签。根据相互关联的量化误差和预测误差共同调整深度网络的学习,最终对给定图像在低维汉明空间上实现快速检索。在MNIST和CIFAR-10两个图像集上的实验结果表明:所提出的方法与几种主流方法相比检索性能更好。最后,将所研究的深度卷积神经网络应用于Metel多媒体教学资源平台的图像检索中,取得了良好的检索效果。

With the explosive growth of network images,the image retrieval method based on traditional artificial design visual features is not very effective. Hence,an efficient supervised hash coding method based on deep convolutional neural network was proposed. The method was based on deep hashing of classification and quantization errors while learning the feature representation of the image,hash code and classifier. Then,the marked supervisory information and the deep architecture were used to learn the convolutional network difference feature representation and directly used to generate the hash code and the predicted image label. Finally,the learning of the deep network was adjusted according to the correlation quantization error and the prediction error,and the fast retrieval of the given image in the low-dimensional Hamming space is realized. Experiments with two image sets in MNIST and CIFAR-10 show that the proposed method performs better than several other mainstream methods. Finally,the proposed deep convolutional network is applied to image retrieval of Metel teaching resources platform and this image retrieval system performs satisfactorily.