A Research Of Unsupervised Hashing Algorithm With Contrastive Learning

Nearest neighbor search has played an important role in recommendation systems and other applications.Its purpose is to accurately find the most similar sample to the query data from a large-scale database.However,it has the problem of dimensionality catastrophe in the case of the rapid increase of high-dimensional data.Then the approximate nearest neighbor search method is proposed and its target is to find the sample that is most likely to be the nearest neighbor.The approximate nearest neighbor search method will bring a certain loss of accuracy,but in search tasks,the approximate nearest neighbor search can not only meet the requirements but also reduce the time complexity.Among them,the hashing method has become one of the widely used methods of approximate nearest neighbor search task due to its high efficiency in calculation and storage.The hashing method converts the original high-dimensional data into a compact binary hash code for storage,then in the subsequent search process,the hash code is directly used to calculate the Hamming distance to find approximate samples.The key of the hashing method is to maintain the consistency between the sample in the original high-dimensional space and the hash code in the Hamming space,so that the original similar data samples have similar binary codes.With the development of deep learning,obtaining hash codes through deep neural network has attracted more and more attention.Fully-supervised deep hashing method has achieved remarkable results,but unsupervised deep hashing method is less effective.In real scenarios,more data samples are unlabeled.When only data samples can be obtained,maintaining the local semantic similarity between samples becomes a difficulty in unsupervised deep hashing algorithms.In addition,most unsupervised hashing methods do not pay attention to how to improve the discriminative ability of the hash code in network learning.This thesis mainly studies the unsupervised deep hashing algorithm,and the specific task is to improve the accuracy of the model to search for images without label information.The method in this thesis is mainly divided into two parts: feature extraction and hash code learning.The stronger feature of sample should be more able to adapt to the negative impact of size,rotation and noise.Therefore,this thesis constructs positive and negative sample pairs by data augmentation,then uses contrastive learning loss to finetune the feature extractor on the target dataset for building a more accurate semantic similarity matrix.Then,in the process of hash code learning,the objective loss function is designed so that the local semantic similarity structure between data in the Hamming space can be maintained,and the discriminative ability of the hash code can be improved by using the idea of contrastive learning.This thesis implements an end-to-end trainable network structure through the Py Torch framework,and conducts experiments on two large-scale image retrieval datasets FLICKR25 K and NUSWIDE to verify the effectiveness of this method.