Research On Image Retrieval Based On Deep Aggregation Features

Image retrieval is a research hotspot in the field of computer vision,and its key technologies mainly include feature extraction and image representation.In recent years,deep learning has been widely used in various fields of society and achieved excellent performance,and its main implementation tool is convolutional neural network.In the field of artificial intelligence,researchers use convolutional neural networks to extract features of images,then perform feature encoding and image representation,and finally use them for image retrieval.Existing studies have shown that deep convolutional features extracted by convolutional neural networks are more conducive to improving the performance of image retrieval.In the deep convolution feature,how to identify the target object and suppress the background noise,and effectively suppress the visual abruptness,is an important problem that needs to be solved urgently.Therefore,this dissertation proposes an image retrieval method based on deep aggregated features and a hybrid PCA-whitening dimensionality reduction method,which greatly enhance the feature representation of the target object,improve the image representation ability,and significantly improve the image retrieval performance.The main innovations of the method proposed in this dissertation are as follows:(1)In order to solve the problem of large amount of background noise and visual suddenness in deep convolutional features,this dissertation proposes an aggregation method to filter deep convolutional features.The method first calculates the variance of the deep convolutional feature maps of each image and sorts the variances,then selects feature maps with larger variance,and uses these feature maps to build filters to filter noise,which can enhance the feature response of the target object,and suppresses background noise.Secondly,the deep convolutional feature maps are weighted by constructing the target space weight and intensity channel weight,which can identify the region where the target object is located and enhance the key features,thereby improving the representation ability of the image and improving the image retrieval performance.(2)In order to solve the problem that the performance of image retrieval is greatly reduced in the process of feature dimensionality reduction,this dissertation proposes a hybrid PCA-whitening dimensionality reduction method.High-dimensional feature vectors require a lot of computing and memory resources.Although the traditional PCA-whitening method can reduce the dimension of the feature,the performance of image retrieval also decreases significantly with the reduction of the feature dimension.Therefore,we propose a hybrid PCA-whitening method,which learns PCA-whitening parameters on two different datasets and reduces the dimension of the features,and then fuses the reduced-dimensional features as the feature representation of the image.It can greatly improve the retrieval accuracy of low-dimensional feature vectors.We compared five well-known image retrieval databases with state-of-the-art image retrieval methods,the experimental results show that the method proposed in this dissertation has excellent retrieval performance,not only can identify target objects and suppress background noise,but also improve the retrieval performance of low-dimensional features.