Research On Image Retrieval Based On Deep Learning

With advances in technology,the development of the Internet and the popularity of handheld terminals,the daily will produce massive digital images.How to find accurate similar images based on user-supplied images in mass images has become a hot issue in recent years.But the traditional use of the characteristics of artificial definition,content-based image retrieval because of the one-sidedness of the feature can't be accurate and effective retrieval of similar images.With the development of deep learning,convolution neural network in image classification showed excellent performance.In this paper,an image retrieval algorithm based on convolutional neural network and isometric mapping algorithm is proposed,which can improve the relevance and accuracy of retrieval.Firstly,this paper deeply studies the traditional content-based image retrieval theory,and researches the basic theory of depth learning domain,and focuses on the analysis and understanding of various convolutional neural networks with superior image classification performance.Finally,the network model of generalized feature extraction for deep convolution neural network is selected.Secondly,based on the application of convolution neural network to image retrieval technology,this paper improves the whole image retrieval process from two aspects.First,the feature extraction process is improved.In particular,by adding a Sigmoid holomorphic hierarchy to the general feature extraction network model of the depth convolution neural network to obtain the binary sequence corresponding to the image feature vector.This operation has the advantage that it is possible to ensure that the new network structure is not degraded in the image classification performance while the binary sequence corresponding to the image can be obtained.The purpose of this approach is to extract similar image feature vector from the image feature vector library according to the binary sequence of the image in the process of image retrieval,and then to further filter similarity according to the feature vector.Second,the image retrieval process is improved.In this paper,we first use the image binary sequence to classify the similar feature set of the image,and then use the isometric mapping algorithm to deal with the eigenvectors.Which is different from the traditional method in that the similarity measure is obtained directly according to the feature vector of the image to obtain the image similar to the input image.It has the advantage of not only can effectively reduce the latitude of eigenvectors,but also can effectively improve the retrieval accuracy.Finally,the Euclidean distance similarity measure is used to retrieve similar images.Finally,the improved retrieval algorithm proposed in this paper is validated by a common image dataset.The results show that the method can effectively improve the retrieval accuracy,and the related images are prioritized.