The Research Of Semantic Image Retrieval Basedon The Deep Convolution Neural Network

Image retrieval problem is a fundamental problem in computer vision,there are lots of scholars have carried out analysis and research about it.At the beginning of this century,people gradually realize the relationship between low-level features and highlevel semantic features inherent in the ”semantic gap”,people began to study the image retrieval based on semantic,but has not made a good progress.In recent years,the development of deep learning technology provides a new tool for solving the problem of image retrieval based on semantic.Compared with traditional method,deep learning model has a stronger ability to learn and it makes a good performance at learning the semantic characteristics of image.Deep learning itself,however,has a lot of research questions,for example,the theoretical analysis of deep learning effectiveness,depth model parameter optimization problems in practical training,and so on.How to design depth for image retrieval based on semantic model also remains further studies.This article studied the development of image retrieval and deep learning technology.Based on this background,a framework of image retrieval based on semantics was put forward.The framework mainly includes three parts:(1)Feature extraction network.It was referred to as a multi-level image semantic feature extraction network in this article,the framework extracts semantic features through the network;(2)The structure of characteristic storage.It will reduce the dimension of extracted features and translate it to a suitable format which can be calculated by image semantic distance measuring formula;(3)Image semantic distance measuring formula.The framework can make the calculation of image semantic distance through this formula.The core of the framework is a network of multi-level image semantic feature extraction,the property of the network influences the accuracy of framework,directly.The network can extract semantic features in the image through its hierarchical structure,the unique structure design helps it to dig deeper semantic image.The features of storage structure turned the extracted features into a fusion feature list at a lower dimension.The image semantic distance measuring formula will calculate the semantic distance through the fusion feature list to finally generate the retrieval result sequence.In the frame of the image retrieval based on semantic,the network of multi-level image semantic feature extraction can be replaced with other classification networks to cope with different retrieval image,flexibility.Test results on multiple data sets show that,compared with existing methods,the framework in this paper can do image retrieval works more accurately based on the semantic.