Research On Hashing Learning Algorithms For Image Retrieval

Data retrieval is the key to various applications in data mining.Especially in the recent years,the explosive growing of the data brings the “curse of the dimensionality” to the data retrieval process and makes the Nearest Neighbor Algorithm unable to meet the actual needs,while Approximate Nearest Neighbor Algorithms are playing an increasingly important role.The hash algorithm has received attention of scholars at home and abroad for its advantages in storage space and computation time.In this thesis,two hash learning algorithms are proposed for image retrieval.Our main work is as follows:The first work of this thesis proposes a random-forest-based hashing algorithm to reduce the space consumption and improve computational efficiency.A random forest which maps the samples in original space into binary hash codes in Hamming space,is constructed.Based on the compact binary hash codes,an order-sensitive Hamming distance is defined to keep the neighborhood of the data in the original space unchanged.Due to the dependency between the feature space and learning algorithms used by different trees,we can flexibly determine the length of hash code in incremental way.In addition,our hashing method based on random forest is naturally suitable for parallel deployment,which can greatly improve the performance of the proposed algorithm.The experimental results of the algorithm are suitable for parallel deployment,which improves the efficiency of the proposed.The experimental results of the proposed algorithm are verified in MNIST and CIFAR-10 data sets.The results shows that the algorithm has high accuracy.The second work of this thesis proposes deep supervised hashing algorithm with N-pair loss function.The algorithm uses the ImageNet dataset to pre-train the network to obtain a deep neural networks that can learn robust image features.Then the hash coding model is trained with N-pair loss.The deep hashing methods based on triplet loss function often suffer from slow convergence and poor local optima,due to that the function employs only one negative sample while not interacting with the other negative classes per each update.Compared with the triplet loss function,the deep hashing methods based on N-pair loss function employs more negative samples per each update,which has the advantage of fast convergence.Finally,fine-grained retrieval is used to improve the accuracy of the algorithm,with a small increase in time.The algorithm is experimentally verified on the MNIST dataset and CIFAR-10 dataset.The results verify the advantages of the algorithm.