Research On Deep Hash Learning For Efficient Image Retrieval

In the big data era,the explosive growth of image and video data has brought unprecedented challenges to efficient inquiry and retrieval of visual targets,Thus there is an urgent need for the study of efficient and robust feature representation and retrieval methods.Therefore,efficient retrieval for large-scale video images has become one of the important problems in computer vision.The key to achieving efficient and robust large-scale image retrieval is how to perform robust compact feature extraction,so the work of this paper mainly focuses on the following three aspects:(1)How to make the back-end compact feature learning and front-end target detection jointly optimize to achieve efficient feature extraction,(2)How to design more robust similarity measurement and quantization methods to improve the discriminability and robustness of compact features.(1)The traditional compact feature extraction methods usually assume that the detection result of the target has been given in advance and only focus on feature extraction modeling.However,in practical applications such as intelligent video surveillance,it is often impossible to assume that the location of the target of interest in the image is known.The performance of the final target retrieval depends not only on the compact feature learning but also on the module of front-end object detection.Facing the problem of large-scale tattoo retrieval,this paper proposes a joint detection and compact features learning method.The proposed approach considers the correlation between the two tasks of object detection and feature learning,so that target detection and feature learning can effectively share features and promote each other.Finally,using an end-to-end learning model to perform the joint optimization of tattoo detection and tattoo compact feature learning simultaneously.(2)For the problem of how to efficiently measure similarity and reduce quantization loss,supervised hash learning method is to perform hash learning using large datasets with the exact label,which usually has higher requirements for image retrieval performance.The existing hash learning methods based on supervised learning are difficult to make full use of the entire Hamming coding space because they usually only consider the discrimination and distance of the hash features in the Hamming space,but rarely consider the hash features The distribution uniformity of Hamming space.This paper proposes a deep hash learning method guided via hamming space anchors.This method uses the pre-division technology of Hamming space and generates multiple sets of structured binary codes,using the structured binary codes as anchors to guide the training of hash learning models;Further,the original feature learning process becomes a process in which the feature points pulled by the anchor points to reduce the quantization loss.Although supervised hash learning has achieved significant progress,in practical application scenarios,large-scale data labeling often requires huge human and material resources.Therefore,in many cases,a large amount of data might be unlabeled.How to use such unsupervised data for effective hash learning is also a very useful but challenging problem.To this end,this paper proposes an unsupervised hash learning method based on generative adversarial networks.This model maximizes the mutual information between the input structured latent variables and the output binary codes through the adversarial learning to achieve the decoupling of intra-class variations and per-class features.The decoupled per-class features are then used to improve the robustness of compact features of unsupervised hash learning.