| With the development of network technology,searching for the required information from the vast Internet images efficiently becomes an important research topic.Traditional content-based image retrieval(CBIR)methods are mainly based on low-level visual features.Since there is a semantic gap between these low-level visual features and high-level human perceptions,traditional methods can not obtain good retrieval result.The study found that deep neural network(DNN)has the ability to abstract high-level semantic features from the data,therefore,more and more research fellows apply deep neural network to the field of CBIR.However,at this stage,the network structure,network training criterion and feature extraction way of DNN are designed for image classification,and may not completely suit for image retrieval.In view of the semantic gap problem and the criterion of feature matching in image retrieval,and based on the network structure and training algorithm of deep neural network,in this paper,we designed a triplet convolutional neural network(CNN)to learn features with the criterion of similarity metric,and extract the non-activation based multi-scale combined features from triplet convolutional neural network,to get suitable high-level semantic features for image retrieval applications.Moreover,referring to the network structure and training algorithm of multi-stage network,this paper designed a triplet multi-stage convolutional neural network,to learn the high-level semantic features,which contain not only global information but also localized information,for content-based image retrieval.The main work is described as follows:(1)Change of network training criterion.The paper optimized the loss function of the network to train the network with the criterion of similarity metric.This paper designed a triplet convolutional neural network based image retrieval algorithm,to learn the high-level semantic features that are more suitable for image retrieval,and applied it to object image retrieval task,texture image retrieval task and fabric image retrieval task.(2)Optimization of the way of feature extraction.This paper proposed to use the non-activation based features taken from deep neural network,to avoid losing any valuable retrieval information during feature activation.And this dissertation also proposed to use the multi-scale combined feature,which is a combination of features from different layers,to make features have more semantic information from different levels and reduce the semantic gap.(3)Solution to the problem of lacking training data.For the applications of image retrieval,this paper used transfer learning to solve the problem of lacking training data.This dissertation proposed to retrain deep neural network on dataset in similar domain,when the retrieved dataset of image retrieval task is not large enough to train the huge network,and apply this training strategy to fabric image retrieval task.(4)Optimization of the network structure.This paper proposed to apply multi-stage network to image retrieval.We design a triplet multi-stage convolutional neural network for content-based image retrieval system by introducing supervision information at the top,middle and bottom of the network.Therefore,the network has the ability to learn the high-quality semantic features,which contain both global information and localized information,and can improve the performance of image retrieval. |
