Image Binary Feature Extraction

For the purpose of filtering out what we care about in a huge amount of visual information quickly,recently large-scale image retrieval has become an important research field for computer vision,in which hash methods,as one of the important techniques of Approximate Nearest Neighbor Search,has been widely concerned for its low s.torage and high speed.Hash methods map visual information from high dimensional space to binary space with a basic requirement of preserving the relative relationships between images after the hash mapping.However,it is expensive to get labeled data and it is difficult for unsupervised hash methods to obtain adequate information.In order to excavate the potential structure of the data,we propose an unsupervised deep hash method and obtain briefer but stronger hash codes.The innovations lie in two loss.functions,firstly,an unsupervised cluster loss assign a pseudo center to each image by analyzing the feature space density,then constantly update the centers meanwhile reduce its number to merge similar subclasses.The second is a ranking loss,which sorts the similarity of samples within a batch for a list-wise fit.The ranking loss stands on a higher perspective and try to directly optimizes the final target.We have evaluated and compared the proposed algorithm with existing methods on CIFAR-10,MNIST and Brown datasets,and found that it has achieved good results in comparison with traditional methods and end-to-end unsupervised deep hash methods,indicating that it makes a very important contribution in retaining the structural information.