Research On Color Image Retrieval Based On Quaternion Representation And Deep Feature

With the development of the internet information era and the popularization of computers and mobile phones,people can upload images of life or entertainment to the network anytime and anywhere,resulting in the image shows explosive growth.Therefore,how to quickly retrieve the needed image from the massive image has become an urgent problem.The general process of image retrieval is extract the image features from the original image database to establish index of the feature database,extract the features of the target image and match them with the features in the feature database,and the match result is taken as the final result of image retrieval.In the existing algorithms of image retrieval,the low-level features or high-level semantic features are usually used in feature matching.There are some low-level feature extraction methods,such as local binary pattern,histogram of oriented gradient,scale invariant feature transformation,etc.Although these methods have achieved good retrieval results,they do not perceive the content of the image,so the retrieval accuracy needs to be improved.There are many methods of high-level semantic feature extraction,among which convolutional neural network correlation algorithm has achieved good results in the field of image.Different features describe different information of the image,and the combination of different features can represent the image more comprehensively.This way of combining features is feature fusion.It is very important to choose the appropriate fusion method during actual feature fusion,which can make the feature matching fast and accurate.This paper will focus on the content-based multi feature fusion of color image retrieval.In this paper,the main works and contributions are proposed three improved image feature extraction methods as follows.(1)Local binary pattern can describe the texture features of local image very well.Due to the widely distribute scope of local binary pattern,the amount of calculation is very large,which is difficult to meet the needs of researchers.In this paper,rotation invariance is added to the local binary pattern,which makes widely distribute scope narrower.According to the local binary pattern with rotation invariance,the second-order full directional derivative calculation method is proposed,and histogram features are extracted.(2)In the color image processing,many existing methods convert color image to gray image,which loses the relationship between color channels.Although many researchers use quaternion to represent color image,these researchers use pure quaternion to represent color image.Therefore,in order to improve the ability of quaternion represent information of image,this paper proposes a method of integrating RGB color channel,information entropy and brightness into the real part and virtual part of quaternion,forming a new matrix and extracting Hu moment features.(3)The image retrieval method based on deep learning has achieved good results in the field of image.Because of the large amount of calculation,it is difficult to apply it to the actual production.In this paper,an image retrieval method based on convolutional neural network is proposed,in which the convolution kernel size of the first two convolution layers of the structure of convolutional neural network is increased,and transforming the first two full connection layers into convolution layers,and then use the improved convolution neural network to extract the deep features with high-level semantic information.The three image features are fused by the weighted method based on the feature dimension,and the Euclidean distance is used to measure the feature similarity.Experiments are carried out on three datasets to verify the effectiveness of the algorithm,and achieved excellent retrieval accuracy with low-dimension features.In addition,the algorithm can also be applied to image classification,face recognition and other fields.