A Image Retrieval Technology Of Complex Concept Based On Deep Learning

With the rapid development of the Internet,there is an explosive growth of images what inspires an urgent need for efficient image search for network users.At the same time,the search styles of users are gradually changing,from a single vocabulary to more complicated sentences.The more complex and longer search sentences have higher requirements for the query technology,not only need to extract the information from the complex query,but also need to find images that match the query in the large-scale image data.Therefore,there is a great practical value to study how to efficiently implement complex query technology for images.In order to solve the above problems,this paper first proposes a weight-based deep learning network model.The model first calculates the concept weight of each concept for the training data set using the frequency of occurrence of the current concept and its related concepts in the training data set.Then,the convolutional neural network(CNN)is used to extract the picture features in the training data set,and the concept classifier of each concept is obtained by combining the above concept weight table.The related concepts of the current concept are determined by the frequency of the two concepts co-occurring in the training data set label.The significance of considering the related concepts is to fully understand the importance of different concepts for the training data set,and it is good for enhancing the accuracy of the concept classifiers.In the test phase,this paper uses the semantic distance-based matching method to map the given complex query text to the corresponding concepts.And it obtains the weight of different concept classifiers at the time of voting by calculating the TF-IDF value of the concept in the label text.Finally,the linear fusion method is used to linearly combine all the concept classifiers and their weights to obtain the total concept classifier,and the total concept classifier judges the picture list matching the given query text.Compared with the traditional average voting method,the weighted classifier voting method adopted in this paper makes full use of the tag information in the training data set,which improves the generalization ability of the network model.In order to verify the effectiveness of the proposed method,this paper do experiments on two data sets separately.The experimental results show that the proposed method can search the images related to the query more accurately.Among the MIRFlickr dataset and the NUS-WIDE-LITE dataset,the proposed method has improved the average accuracy by3.7% and 4.1% respectively.