基于深度学习哈希算法的快速图像检索研究

Fast Image Retrieval Based on Hash Algorithm in Depth Learning

现在存在的大部分监督哈希是将手工提取的特征转换为哈希值,然后根据图像标签为监督信息得到损失函数,但是手工提取特征以及不完全考虑所有损失的损失函数会降低检索精度。监督哈希算法主要目的是通过训练数据以及数据的标签提升数据与相应哈希的相似度,从而提高检索的相似度。论文提出了一个新的监督哈希算法,将每个图像的多标签转换为二进制向量,通过汉明距离得到成对图像的相似度,放入损失函数中作为监督信息,加上图像特征量化为哈希码时的量化误差以及所有图像哈希码与平衡值的差值,结合以上所有部分生成损失函数,进行网络训练。实验结果显示论文的方法在检索精度上比现有的方法有所提升。

Most of the existing surveillance hash is converting the hand extracted features into hash values,and then the loss function is obtained based on the image labels for supervised information,but manually extracting features and losing all loss function will reduce retrieval accuracy.The main purpose of the supervised hash algorithm is to improve the similarity between the data and the corresponding hash through the training of data and the labels of the data,so as to improve the similarity of the retrieval.This paper proposes a new supervised hash algorithm,the multi label of each image is converted to binary vectors obtained by pairwise similarity image Hamming distance into the loss function as supervised information,image feature quantization and quantization error for hash code and image difference balance value,combined with the above all parts of production loss function,training network.The experimental results show that the method in this paper is better than the existing method in the retrieval precision.