一种多尺度平衡深度哈希图像检索方法

Multi-scale balanced deep hashing method for image retrieval

传统监督哈希方法将图像学习的手工特征或机器学习特征和二进制码的单独量化步骤分开,并未很好地控制量化误差,并且不能保证生成哈希码的平衡性。为了解决这个问题,提出了新的多尺度平衡深度哈希方法。该方法采用多尺度输入,这样做有效地提升了网络对图像特征的学习效果;提出了新的损失函数,在很好地保留语义相似性的前提下,考虑了量化误差以及哈希码平衡性,以生成更优质的哈希码。该方法在CIFAR-10以及Flickr数据集上的最佳检索结果较当今先进方法分别提高了5. 5%和3. 1%的检索精度。

The use of the semantic similarity improving the hash coding quality has recently been more widely concerned. Traditional supervised hash methods for image retrieval represent an image as a manual feature vector or a machine learning feature vector,and then perform a separate quantization step to generate a binary code. Such methods do not control the quantization error effectively,and cannot guarantee the balance of hash code. To this end,this paper presented a new multi-scale balanced deep hash method. The method used multi-scale input,which effectively improved the ability of learning the image features from the network. Moreover,it proposed a new loss function. Under the premise of preserving the semantic similarity,it took the quantization error and the balance of hash code into account to generate the high quality hash code. After experimenting on two benchmark databases: CIFAR-10 and Flickr,this method has been improved by 5. 5% and 3. 1% of the search accuracy compared with today’s advanced image retrieval methods.