Research On Efficient Image Retrieval Based On Deep Hash Learning

The focus of the visual retrieval field is how to accurately and quickly from the image database retrieval images related to user queries.With the explosive growth of image data,text-based image retrieval has gradually evolved into content-based image retrieval due to the increased workload of annotations and the subjectivity of annotation content.In content-based image retrieval,there are two key issues to be studied.One is to study feature extraction algorithms that can better represent image content and image semantics.The other is to study how to quickly match the image features of the query to the similar features in the massive image feature database.Based on the above key research points,researchers have proposed deep hash learning methods in combination with deep convolutional neural networks and hashing techniques in recent years.The method greatly improves image retrieval accuracy and speed.However,there are some problems in the proposed method.For example,CNNH cannot extract features and represent hash features at the same time.DNNH needs to select triplet images before training the model.SSDH needs to manually set the learning rate when training the network,and the training speed is slow,and in the retrieval process.In addition,the deep hash learning algorithm rarely considers the establishment of hash feature index to improve query efficiency.In view of the above shortcomings,this thesis designs an efficient image retrieval algorithm,the main contributions are as follows:1.The model for extracting hash features is studied.A hash layer is added after the fully connected layer in the model,and a low-dimensional feature is obtained by reducing the number of nodes of the hash layer.Moreover,the designed model has fewer network parameters and is easy to train.The Ada Delta algorithm is used in the model training,which avoids the manual setting of learning rate and speeds up the model training.Once the model is trained,the binary hash feature of Hamming space is obtained by the threshold function.The analysis shows that the image hash feature learned by the model can represent the high-level semantics of the image,and has uniformity and intra-class similarity.2.Constructing an index structure based on hash feature.In the face of large-scale binary hashing,a hash feature index needs to be established in order to improve query efficiency.In the thesis,an index structure is designed based on the uniformity of hash features and intra-class similarity.The index structure combines the partitioning idea and the inverted index.Through the query algorithm and theoretical analysis,it can be concluded that the indexing can not only speed up the image retrieval,but also ensure that the retrieval precision under the index is basically consistent with the linear retrieval precision.Experiments were carried out on the proposed hash feature extraction algorithm and hash feature index structure on the public dataset commonly used in image retrieval.The accuracy evaluation index MAP indicates that the hash feature extraction algorithm proposed in the thesis is better than other classical algorithms in the Cifar10,Mnist and Oxford17 datasets.When the appropriate parameters are selected,the efficiency evaluation index ART indicates that the index improves the retrieval efficiency by 5 times on the Cifar10 and Mnist datasets,and the retrieval efficiency of the millions of dataset Image Net has been improved by 20 times.In addition,under the index we proposed,retrieval precision MAP is basically the same as the MAP of the linear retrieval.