Research On Deep Hashing Network For Large Scale Image Retrieval

Image retrieval plays an important role in machine learning,artificial intelligence and computer vision.However,with the explosive growth of data and the increase of data dimensions,the large-scale image data brings new difficulties to image retrieval technology in storage and calculation.In recent years,the image retrieval algorithm based on hash computation has been widely studied in the field of computer vision due to its low storage cost and fast retrieval speed.In the image retrieval algorithm based on hashing calculation,there is a key problem that is how to effectively process the high-dimensional feature vectors and optimize the hash function to improve the accuracy of the retrieval algorithm.Aiming at this problem,we carry out our research work in this thesis.We combine the deep neural network with the hash algorithm for adaptive learning,which can not only deal with data sets with and without labels,so it has better applicability.The framework of the algorithm is divided into three parts:first,the complex convolutional neural network is used to generate hash tags;secondly,the obtained hash tag is used to train the hash function;thirdly,image retrieval based on the acquired training model.We have made relevant improvements in parts(2)and(3),including the following:1.In order to optimize the training hash function and ensure the output more stable when calculating the loss function,batch normalization algorithm is generally used to reduce the deviation.However,in order to further improve the accuracy of the algorithm,three normalization methods are introduced,namely,layer normalization,instance normalization and group normalization.The experimental results show that the performance of group normalization is the best.Four dimensionality reduction methods,including principal component analysis,Laplace characteristic map,local linear embedding,and independent component analysis,are used to train the high-dimensional feature vectors of the full connection layer of hash function.They realize the dimensionality reduction processing of high-dimensional features and improve the accuracy and speed of the algorithm.2.In order to make the output of the hashing layer neurons closer to 0 and 1,a more uniform binary hash encoding is obtained to reflect the class differences between different images.An improved loss function is proposed,that is,a penalty term is added to the original loss function.However,in order not to make the first penalty term extreme,which is all zero or one,leading to a big deviation in the hashing,we added a second penalty term to constrain it.The experimental results also prove the effectiveness of the improved loss function.3.In order to further improve the accuracy of image retrieval,we propose a two-level retrieval framework.First,the ranking of image similarity is obtained through first-level retrieval in Hamming space.The top 20 images are selected.Their original image feature vectors and image feature vectors of query images are calculated one by one by Euclidian distance.The similarity is ranked from high to low and as the final retrieval result.Experiments in public image sets Cifar10,Cifar100(fine),Cifar100(coarse),and STL-10 show that the algorithm proposed in this thesis is better than other hashing algorithms in terms of various metrics.