Image Hashing Retrieval Based On Auto-Encoder

Hashing learning methods play an important role in the field of image retrieval.They are efficient and fast and occupy relatively small storage space in processing large-scale data,so they have excellent retrieval performance.Among them,the hashing retrieval algorithm based on binary auto-encoder is applied in the field of image retrieval.By minimizing the reconstruction loss,this algorithm makes the hash codes mapped to the Hamming space keep the important information of the original input data as much as possible.However,the hashing retrieval algorithm based on binary auto-encoder only considers the reconstruction loss,which is not good for learning high-quality hash functions.On the one hand,since the algorithm doesn't consider the local geometric structure information between the original input data,the learned hash codes can't well represent the original data.On the other hand,since this algorithm is an unsupervised method,and it can't make full use of semantic labels information in the process of alternately optimizing the model,which is unfavourable for generating hash codes with discriminative ability.Therefore,by adding additional constraints to guide the learning process of the hashing retrieval model based on binary auto-encoder,this thesis proposes auto-encoder image hashing retrieval algorithm based on manifold similarity preservation and auto-encoder image hashing retrieval algorithm based on semantic similarity preservation.The research work of this thesis is as follows:1.For the problem that the hashing retrieval algorithm based on binary auto-encoder doesn't pay attention to the local geometric structure information of the original input data,this thesis proposes auto-encoder image hashing retrieval algorithm based on manifold similarity preservation(MSPAH).This algorithm combines with supervised Laplacian eigenmaps algorithm for the generation of the referenced hash codes to implement local invariance constraint into the hashing retrieval model based on binary auto-encoder.First,the supervised Laplacian eigenmaps algorithm for the generation of the referenced hash codes is used to generate referenced hash codes,which keep the local geometric structure information of the original input data.Then by using the hash codes as a reference,the manifold similarity preserving loss is constructed between the hash codes generated by the encoder and the referenced hash codes to guide the learning process of the model,so that the MSPAH model can provide strong characterization ability while keeping the local geometric structure information unchanged as much as possible,so as to learn hash functions with stronger generalization ability.Experiments on three benchmark datasets,i.e.,CIFAR-10,MNIST and NUS-WIDE,show that this algorithm can effectively improve the performance of image retrieval.For CIFAR-10 dataset,when the number of returned samples is 1000 and the code length is 16 bits,the precision of MSPAH is about 3.7% higher than that of the hashing retrieval algorithm based on binary auto-encoder.2.For the problem that the hashing retrieval algorithm based on binary auto-encoder doesn't make full use of semantic labels information in the process of alternately optimizing the model,this thesis further improves the proposed auto-encoder image hashing retrieval algorithm based on manifold similarity preservation,and proposes auto-encoder image hashing retrieval algorithm based on semantic similarity preservation(SSPAH).By regressing the semantic labels to the Hamming space,this algorithm constructs a semantic similarity preserving loss between the hash codes generated by the encoder and the corresponding hash codes of the semantic labels matrix,making the hash codes generated by the encoder more discriminative while maintaining the local geometric structure information of the original input data,and further improving the retrieval performance of the model.The experimental results show that the retrieval accuracy of this method is better than that of MSPAH and several other existing comparison methods,which verifies the effectiveness of integrating semantic labels information into the auto-encoder model.For NUS-WIDE dataset,when the number of returned samples is 1000 and the code length is16 bits,the SSPAH algorithm is about 1.9% better than MSPAH and 5% better than the hashing retrieval algorithm based on binary auto-encoder.3.Design and implement an image hashing retrieval simulation system based on autoencoder.The system integrates the two hashing retrieval algorithms proposed in this thesis,as well as RBA,BA,ITQ,KMH and SPH five hashing retrieval comparison algorithms.The system has strong flexibility.Users can select different retrieval models according to their needs through a simple and easy-to-operate interface for comparison of retrieval results.After importing the image to be queried and setting retrieval parameters,the system will display the retrieval results of the selected retrieval models according to the needs of users,so as to facilitate the comparison of retrieval performance for users.