Research On Image Retrieval Based On Deep Convolutional Neural Network

With the development of Internet technology and the popularity of smart devices,image data uploaded to the network is increasing.Faced with a large amount of image data,how to accurately and efficiently retrieve the required images is a research hotspot and difficulty in image retrieval tasks.Content-based image retrieval technology relies on extracted description features for retrieval,so the ability to describe image features largely determines the accuracy of retrieval.In recent years,deep convolutional neural networks have made some progress in image retrieval tasks by virtue of their powerful feature extraction capabilities.However,the existing deep convolution features have insufficient description capabilities,and the accuracy of retrieval needs to be improved.And,there is a problem that the extracted high-dimensional image features cause a decrease in retrieval speed.Based on the above problems,this paper explores how to extract more discriminative features from deep convolutional neural networks and reduce the feature dimensions used for retrieval.The main work and results are as follows:1)For the problem of insufficient description of the local details of the fully connected layer features extracted by the network,a multi-region center based convolution feature weighting aggregation method is proposed.The method relies on the perceptual ability of the convolutional layer in the deep convolutional network model to the semantic region of the image.Using the difference in the response value of the convolutional feature map,the feature map containing the information of the target region is selected,according to the higher convolution in the selected feature map.The response center calculates the regional center weight,and finally weights the convolution activation response to generate an image global description.In this paper,the experimental comparison of the commonly used landmark image datasets is carried out.The experimental results show that the retrieval accuracy is improved compared with other mainstream convolutional feature weighting aggregation methods.Compared with the suboptimal algorithms in Paris architecture and Oxford building datasets,The average search accuracy averaged by 1% and 5%,respectively.2)For the problem that the network model using tuple training does not fully exploit the ranking information contained in the sample,a depth binarized hash coding feature extraction method based on ranking weighted triples is proposed.The method utilizes the image triplet containing the misaligned information to calculate the corresponding tuple weight,and uses the improved target loss function to guide the update of the network layer parameters,and the extracted binarized features still maintain similar images in the Hamming space.The similarity between the two dimensions and the high-dimensional features are encoded into a compact binary encoding,which reduces the feature dimension and improves the retrieval speed.The experimental results show that the proposed method can effectively improve the retrieval performance of the network extraction features.The accuracy of the CIFAR-10 test is 1% higher than that of the suboptimal method,and the average time consumption of the single search is reduced by 0.25 ms,while the NUS-WIDE data is reduced.The set increased by 2% and the time consumption decreased by 0.16 ms.