Based on Lightweight Network and Multiple Supervised Hash for Image Retrieval
Zhuang Chen, Jiahao Qiu, Yuan Zhao · 2022 IEEE International Conference on Artificial Intelligence and Computer Applications (ICAICA) · 2022
At present, most of the depth hash methods use complex convolution neural networks to extract the features of the image, many parameters and large models, and insufficient utilization of feature correlation. A method called LNMSH (lightweight network and multi-supervised hash for Image Retrieval) is proposed based on lightweight network and multiple supervised hash. LNMSH is trained as image inputs to learn hash codes through a designed lightweight convolution neural network and a multiple supervised hash function. A lightweight convolution neural network has few parameters and a small model to efficiently extract the image feature information. The multiple supervised hash function can more efficient utilization of the semantic information and the feature correlation between images, improving the conversion of the feature information extracted by the network to a more distinguished hash code and improving the efficiency of image retrieval. Therefore, a total target loss function is designed. Many comparative experiments have been performed on two widely used datasets, ImageNet and Nus-Wide. Verified by the experimental results, the performance of LNMSH is better than that of other deep hash algorithms.