Study On Semantic-aware Ranking Preserving Deep Hashing Image Retrieval

The image retrieval has important application value and research significance.At the same time,it also faces a series of practical challenges including how to improve the retrieval speed and generate accurate ranking results.The hashing technology can effectively improve the speed of distance calculation by mapping features from high-dimensional space to low-dimensional binary space,which has become an important method to improve the retrieval speed.In addition,deep learning can alleviate the "semantic gap" problem between the low-level visual information and high-level semantic information of images to some extent.Thus,the combination of deep learning and hashing methods for image retrieval has important research value and significance.This thesis aims to focus on the multi-label image retrieval based on deep learning and hashing method.Aiming at the current three research problems in the field of multi-label retrieval,in-depth research is conducted and the corresponding solutions is proposed.The main innovations and contents of this thesis are as follows:(1)Aiming at the problem of insufficient semantic expression ability of similarity quantization methods for multi-label images,a novel quantization method is proposed which combines category and semantic information.Different from the methods that only consider the category level similarity,this thesis uses natural language processing technology to encode the image captions and further injects them into the quantization method as auxiliary information,which can capture high-level semantic information beyond the category and improve the semantic expression ability of the quantization method.(2)Aiming at the problem that general pairwise loss functions are insensitive to the relative order of similar images,the upper and lower threshold functions that can be adaptively changed with the similarity matrix are constructed.By integrating the upper and lower threshold functions into the original pairwise loss function,a compact zero loss distance interval is constructed in the Hamming space,which can directly constrain the relative order of similar images.The physical meaning of this constraint is to encourage images with higher similarity to be ranked ahead in the retrieval results.(3)Aiming at the data imbalance problem during the training stage,the adaptive weight coefficient are assigned into the pairwise loss function according to the amout of similar and dissimilar image pairs in a training batch,which can adaptively adjust the contribution rate to the total loss and reduce the impact of uneven distribution of dataset during training.(4)Combining the above three points,this thesis proposes a deep semantic-aware ranking preserving hashing(DSRPH),and further constructs a neural network framework for multi-label image retrieval.DSRPH has a stronger perception of high-level semantic information and can generate more accurate retrieval ranking result based on multi-level similarity.Through the multi-angle experimental evaluation on the three multi-label image public datasets of MS COCO,NUS-WIDE and Flickr,the validity and the superiority of DSRPH in retrieval performance is verified,compared to the six representative excellent hash methods of IDHN,Hash Net,DHN,DSH,ITQ and SH.