Research On Image Retrieval Method Based On Convolutional Neural Network And Hash Technology

With the rapid development of computer information technology and the application and popularization of multimedia processing equipment,the data volume of digital images has shown an exponential and rapid growth trend.Facing the increasing digital image resources,how to accurately find the image data required by the system administrator from a large-scale image data set has become an urgent technical problem to be solved.The text-based image retrieval method was first applied,and then the content-based image retrieval method effectively solved the defects of large workload and irregular expression of artificial semantic labels by using characteristics such as color,texture,and spatial shape.However,due to the high dimension of the image features extracted by this method and the inability to fully express deep semantic features,the storage consumption has increased and the retrieval speed and accuracy have dropped rapidly as the amount of image data increases.In order to accurately represent the deep semantic features of the image and solve the problem of high-dimensional feature data,this paper proposes an image retrieval method based on convolutional neural network and hash technology.The method first combines the powerful deep learning capabilities of the convolutional neural network model to extract image features,obtain the implicit connections between visual features of similar images,and obtain deep semantic feature data of digital images;then through a pre-trained stacked autoencoder Reduce the dimensionality of the feature vector,use the sparsity of the feature data and Frobenius norm to constrain the stacked autoencoder,and increase the robustness of the algorithm by adding regular penalty items;then use iterative quantization hash learning method to adjust the rotation The matrix minimizes the quantization error of the mapping to obtain the binary hash code of the image feature map.Finally,the Hamming distance is used to calculate the similarity between the images.In order to avoid the adverse effects of the same Hamming distance,the weighting optimization method is used to adjust Arrange the order to get the final search result set.The experimental results show that the test is performed on Cifar-10 and Caltech-256 public image retrieval dataset.This method can combine the learning advantages of the deep network model to extract short and efficient deep image semantic features,reduce the impact of the high dimensionality of the feature vector,and effectively improve the accuracy and speed of the image retrieval system.