Research On Large-scale Image Retrieval Based On Heterogeneous Feature Fusion

With the rapid development of the Internet and the widespread popularity of mobile photographic equipment,the number of images and the diversity of image content is becoming increasingly rich,and gradually showing a trend of explosive growth.Faced with such a large amount of complex image data,how to quickly and effectively find the information one needed is becoming a hot topic of current research.We know that a picture is worth a thousand words,so scholars have focused on the content-based image retrieval technology.The content-based image retrieval process mainly includes three stages: the extraction of image visual features,the construction of feature indexes and the formulation of similarity matching strategies.The discrimination of the extracted image visual information is critical to the accuracy of the final retrieval.Unfortunately,until the current image retrieval technology,there is no feature extraction algorithm that can accurately express all the content contained in the image,so the study of multi-feature effective fusion search method is particularly necessary.However,how to ensure the feature dimension and the retrieval efficiency become another problem,for example,the parallel fusion of multiple image visual features requires the storage of multiple visual features on the database image,resulting in an increase in storage consumption.This thesis presents the shortcomings of existing image retrieval models,and absorbs the latest achievements in deep learning technology.On this basis,a large-scale image retrieval framework based on heterogeneous feature fusion is proposed and named as the adaptive fusion framework.In particular,this framework aims to realize the fusion of global features based on the convolution neural network and local features based on the manual design,and then design two corresponding modules named global retrieval and accurate query for the above two characteristics.The global retrieval module mainly completes the filtering operation,that is,realizes the fast retrieval of the dataset through global features,filters out noisy images that are not related to the query image,and learns a weight for each candidate image to measure its similarity degree to the query image.The purpose of the exact query module is to complete the optimization operation in the candidate images,that is,accomplishes further retrieve of candidate set through the improved BoW model,and fuses its search results with similar weights obtained via global retrieval module to get the final similarity values.The adaptive fusion framework takes the two-layer retrieval mechanism as the core,and uses the filtering and optimization operations to ensure the full efficiency and adaptability of the retrieval.By comparing the results of multiple datasets,compared with current methods,the adaptive fusion framework proposed in this thesis can take full account of the efficiency and accuracy of image retrieval tasks.