Image Retrieval Algorithm Of Research Based On Deep Hash

With the rapid development of artificial intelligence and the advent of the 5G era,people enjoy the convenience brought by the network,and a large amount of image data also grows exponentially.How to get the desired image accurately and quickly becomes a challenging subject in image retrieval.The hash algorithm expresses high-dimensional feature vectors as a compact binary code,and uses Hamming distance sorting to query the nearest neighbors.It has the characteristics of less storage space and simple calculation,and has become a research hotspot.The traditional hashing algorithms make the retrieval efficiency inefficient because of its weak image expression ability.In view of the excellent performance of the deep hash algorithm in improving image retrieval performance,this thesisf ocuses on deep research based on the deep hash algorithm for image retrieval.The main work of the thesis is as follows:1.Two deep hash algorithms combining multiple loss functions are proposed.In the optimization stage,in order to increase the supervision information and solve the multilabel classification problem at the same time,a deep hash algorithm based on classification loss function is proposed.For the problem of gradient disappearance caused by the output value of the activation function,a deep hash algorithm based on constrained loss function is proposed.First,the features of the image are extracted through the convolutional neural network,the feature vector of the image is output through the fully connected layer,and the binary hash code is generated in conjunction with the product quantization method.Finally,the proposed classification loss function and constraint loss function are respectively combined with the edge loss function and quantization,the three loss functions are combined into the target loss function.The target loss function is minimized by the Adam optimizer,thereby improving the image retrieval performance.2.A deep hashing algorithm combined with differentiation module is proposed.For the deep hash algorithm based on the constraint loss function proposed in this thesis,there is still a problem of redundant information between feature bits.Based on the original framework,a differentiation module is introduced in the feature extraction stage to convert the fully connected layer into a locally connected module,each quantization bit position is only related to part of the input,and then combined with the product quantization method to encode,the final output is a more compact binary encoding.Experiment on three datasets of CIFAR-10,NUS-WIDE and Image Net.The improved deep hash algorithm proposed in this thesis is compared with the commonly used image retrieval algorithms.The experimental results show that the algorithm proposed in this thesis can effectively enhance the expression ability of image features and improve the performance of image retrieval.