基于深度哈希的批量图像并行检索方法

Batch Images Parallel Retrieval Based on Deep Hashing

针对图像检索的精确度和效率基于内容海量图像检索的关键问题,提出了一种基于深度哈希算法的图像并行检索方法。首先使用卷积神经网络建立图像特征和哈希码提取模型,然后将图像输入到训练好的模型中获取图像特征和哈希码,并存储在分布式数据库HBase中,最后在Hadoop并行计算框架中实现了一种并行检索方法。在大规模数据集CIFAR-10上进行检索实验,得到平均准确率为60.28%,相比SIFT算法提高了12.63%,且批量检索一张图像的平均时间为0.73 s。因此,该方法可使检索精度得到明显提高,还能提高海量图像的存储和检索效率。

In view of the accuracy and efficiency of image retrieval is the key problem in contentbased image retrieval, a batch images distributed retrieval algorithm based on deep hashing is proposed. Firstly,the image features and hash code extraction model is built by training deep hashing model using the training data. Then,the features and hash codes of images are extracted by the model and stored in the distributed database HBase. Finally,a parallel retrieval method is implemented with Hadoopframework. The experiment on a large scale dataset CIFAR-10 shows that the method is effective. The mean average precision is 60. 28%. The accuracy has increased by 12. 63% comparing to SIFT,and the average retrieval of an image takes 0. 73 s. It can not only solve the problem of massive images storage and fast retrieval,but also improve the accuracy.