Image Retrieval Research Based On Deep Learning

Since 2012,the convolutional neural network has achieved great success in the field of image recognition.Many researchers began to use pre-trained CNN model as feature extraction in a variety of computer vision tasks,such as object detection,object recognition,image retrieval and so on.Image retrieval task not only need to pay attention to class represented by the gap between classes,but also need to consider the difference within the same category to distinguish between them.But when CNN global features are used in image retrieval task,it only focused on the global semantic information.They are lack of the local details of image and can't effectively define which image contains multiple objects.So the retrieval result is not ideal.This thesis studies the local feature extraction algorithm based on deep learning.We use the aggregation strategy and significant regional strategy to generate local feature which can be used in image retrieval.The main work is as follows:Firstly,we introduce the theoretical foundation of deep learning and briefly introduce several kinds of traditional image retrieval algorithm.Then,we introduce several classic deep learning models.In addtion,the current study on image retrieval related work using deep learning are analyzed and summarized.Secondly,through analyzing the defect of the global CNN which cannot be effectively applied to image retrieval task to describe the local details,this thesis proposes a polymerization CNN low-level feature maps to generate local characteristics.CNN high-level features of the model have more semantic information,but the underlying characteristics pay more attention to local details.The higher layer,the more abstract feature.Based on this nature,we extract low-level CNN feature maps,and through the channel weighted aggregation and space weighted aggregation and BOW aggregation,we generate local features to describe the local image details.Compared to the global CNN features,applying the local characteristics to the same object image retrieval task,the retrieval accuracy effectively enhances.Thirdly,by using the traditional local feature descriptor generation process,we generate the deep local features through three step : significant areas,significant area description and coding.The traditional local descriptor still has great advantages in image retrieval.Therefore,in this thesis,we use significant area instead of the the key point in traditional local features to generate deep local features.We use the method of image understanding to extract significant areas,and has obtained the good retrieval effect in image retrieval task.We conduct simulation experiment in the classic image datases,and compare with the excellent image retrieval algorithms.Experimental results show that the effectiveness of the method,which can further enhance the accuracy of image retrieval and has strong practical applicability.