A deep locality-sensitive hashing approach for achieving optimal ‎image retrieval satisfaction

Efficient methods that enable high and rapid image retrieval are continuously needed, especially with the large mass of images that are generated from different sectors and domains like business, communication media, and entertainment. Recently, deep neural networks are extensively proved higher-performing models compared to other traditional models. Besides, combining hashing methods with a deep learning architecture improves the image retrieval time and accuracy. In this paper, we propose a novel image retrieval method that employs locality-sensitive hashing with convolutional neural networks (CNN) to extract different types of features from different model layers. The aim of this hybrid framework is focusing on both the high-level information that provides semantic content and the low-level information that provides visual content of the images. Hash tables are constructed from the extracted features and trained to achieve fast image retrieval. To verify the effectiveness of the proposed framework, a variety of experiments and computational performance analysis are carried out on the CIFRA-10 and NUS-WIDE datasets. The experimental results show that the proposed method surpasses most existing hash-based image retrieval methods.