Research On Image Retrieval Algorithms Based On Deep Neural Networks

Content-based image retrieval(CBIR)is a hot point in the field of computer vision.With the rapid development of the convolution neural networks(CNN)in recent years,their applications in the CBIR systems have becoming more and more popular.However,in most existing CBIR systems,studies were only focused on building a better CNN network model to extract more abstract image features or finding better feature matching algorithms to improve retrieval performance and the impact of the information contained in the feature space and the semantic of the image on the retrieval system were usually neglected.To solve the above two problems,the main work and contributions of this thesis are as follows:Firstly,a retrieval method based on the convolution neural network and Efficient Manifold Ranking(EMR)is proposed.In this method,considering the fact that the dimension of CNN feature is usually very high,the Kernel Principal Component Analysis(KPCA)algorithm is first utilized to reduce the dimension of the feature.Then Efficient Manifold Ranking method is used to dig out the neighborhood information containing in the feature space.Finally,for a given query image,the feature points are sorted according to the neighborhood information between these feature points and the query feature points,and the image corresponding to the higher score is considered to be relevant to the query.Experimental results show that the algorithm proposed in this thesis is superior to the popular hash methods.Secondly,an image retrieval algorithm based on eye-tracking data and relevance feedback(RF)is proposed.Firstly,in this paper,support vector machine(SVM)and fuzzy C-means clustering(FCM)algorithm are used to classify and pre-filter the images to reduce the search space and improve the retrieval efficiency.Secondly,to narrow down the semantic gap,this paper utilized the eye tracking technique and the eye tracking data as away for feedback information acquisition.Experimental results demonstrate that the combination of eye-moving data and relevance feedback can significantly improve the performance of the retrieval.system.