Multi-Grained Cascade Forest Based Hashing For Image Retrieval

Hashing methods have been intensively studied and widely applied in the field of content-based image retrieval.Hashing methods aim to learn a group of hash functions to map the original data into binary codes and simultaneously preserve some notion of similarity in the Hamming space.The generated binary codes can improve the efficiency of retrieval and greatly reduce the storage cost.The decision tree is a fast and interpretable model,but the current decision tree-based hashing methods have insufficient learning ability due to the use of shallow decision trees.Most of the current deep hashing methods are based on deep neural networks.However,considering the deficiencies of deep neural network-based hashing,such as the presence of too many hyperparameters,poor interpretability,and requirement for expensive and powerful computational facilities during the training process,a non-deep neural network-based hashing model need to be designed to achieve efficient image retrieval with few hyperparameters,easy theoretical analysis and an efficient training process.The multi-grained cascade forest(gcForest)is a novel deep model that generates a deep forest ensemble classifier to process data layerby-layer with multi-grained scanning and a cascade forest.To date,gcForest has not been used to generate compact binary codes;therefore,this thesis proposes a multi-grained cascade forest based hashing method for image retrieval and a multi-grained cascade forest based hashing method with manifold similarity preserved for image retrieval.In order to further reduce the storage cost of the image database,this thesis proposes a random forest autoencoder-based hashing method for image retrieval.The specific research work is as follows:1.The existing decision tree-based hashing methods use shallow decision trees,which may lead to insufficient learning ability.And the deep neural network basedhashing methods have some deficiencies like too many hyperparameters,poor interpretability and the requirement of expensive computing device and large-scale datasets during the training process.To solve these problems,this thesis proposes a multigrained cascade forest-based hashing method for image retrieval.The method first scans original data using a multi-grained sliding window to extract multi-grained features,and then uses a two-step learning strategy,i.e.,initial binary code inference and deep forest hash function learning to realize hash mapping.The experimental results show that the method is simple in parameter setting,and the retrieval accuracy is higher than that of the decision tree-based hashing method and deep neural network-based hashing method.2.In order to fully consider the semantic similarity and attributes of data,this thesis proposes a multi-grained cascade forest-based hashing method with manifold similarity preserved for image retrieval.This method considers semantic similarity and manifold similarity when constructing the formulation of the loss function to map the data from the original space to the Hamming space preserving the both semantic similarity and manifold similarity of the data in the original space.In image retrieval tasks,the related image that is simultaneously preserved in the label semantics and the content semantics is retrieved based on the content information of the query image,thereby the images that is closer to the semantic information of the query image can be retrieved.The experimental results show that the retrieval accuracy of this method is higher than the decision tree-based hashing method and deep neural network-based hashing method.Furthermore,users can obtain the images that are more close with query image in semantic.3.In order to further reduce the storage cost of image database in image retrieval tasks,this thesis proposes a random forest autoencoder-based hashing method for image retrieval.The method uses random forest autoencoder to decompose the image retrieval task into forward encoding process and backward decoding process.The forward encoding process maps data from image space to Hamming space,and the backward decoding process decodes the hash codes from Hamming space to image space.Therefore,in the image retrieval system after the completion of the encoding,it is no longer necessary to save the image database,and the images can be directly reconstructed by using the hash codes,thereby greatly reducing the storage cost.The experimental results show that the method has a lower retrieval accuracy,but it can reconstruct the image well.