Deep Hashing With Hadamard Matrix For Image Retrieval

The amount of image and video data in social networks and search engines are growing at an alarming rate.In order to effectively retrieve large-scale high dimensional image data,Approximate Nearest Neighbor search has become a hot research topic.Among various Approximate Nearest Neighbor technologies,hashing has become a popular solution due to its low storage cost and fast retrieval speed.Traditional hashing is based on hand-crafted features,but these features can only extract the underlying representation of data,which may fail to maintain the semantic similarity of image pairs.Meanwhile,the learning process of hash functions is independent of feature extraction step,which makes hash code learning unable to give any feedback to feature learning step.In recent years,deep learning has been successfully applied to hash problems.Deep hashing can utilize deep neural network to learn features and hash codes simultaneously,thus the performance of deep hashing is usually superior to traditional hashing.At present,most deep supervised hashing can not fully learn the similarity relationship between data pairs and lack the guidance of quantitative statistical information,which leads to limited performance.Besides,these deep supervised hashing usually impose artificial constraints of bit independence and bit balance on hash functions,which potentially restrict the flexibility of hash functions to fit training data,resulting in complicated optimization problems.To solve the problems mentioned above,we propose a Hadamard Matrix Guided Deep Cauchy Hashing(HMDCH),which can generate high-quality hash codes in an end-to-end deep architecture.Specifically,each column of Hadamard matrix is regarded as the hash center corresponding to different image category,which essentially satisfies the requirement of bit independence and bit balance and can serve as the desired outputs of hash functions learning.On this basis,a probability function of Cauchy distribution is combined to measure the logarithm posterior estimation of hash codes,and a Cauchy central similarity loss is further proposed to significantly punish all images whose Hamming distance between the generated hash codes in network and the corresponding hash centers are larger than given Hamming radius threshold,thus encouraging hash codes of similar data pairs to approach a common center,and those for dissimilar data pairs to converge to different centers.We also introduce Cauchy quantization loss to make hash codes converge to hash centers completely.Extensive experiments on multiple widely-used datasets demonstrate that HMDCH can generate highly centralized,discriminative,and balanced binary codes and improves the performance of image retrieval.In addition,an image retrieval system based on HMDCH is designed and implemented by combining various development technologies.