Research On Academic Image Retrieval Methods Based On Multi-feature Fusion

Academic literature retrieval is an important part of academic research.At present,the academic literature retrieval is still mainly based on the text-based retrieval method,which relies on the title,key words and other information of the academic literature,and has the problem of subjectivity and ambiguity.However,most academic literature is accompanied by a large number of images.Therefore,the research on academic image retrieval methods will effectively promote the progress of academic literature retrieval technology.Focusing on how to improve the accuracy,efficiency and self-adaptability of academic image retrieval methods,this thesis proposes an image retrieval method based on the enhanced rotation invariant LBP feature extraction and the multi-feature fusion image retrieval method based on adaptive weight.The method effectively improves the accuracy,efficiency and self-adaptability of academic image retrieval.The research focus of this thesis is as follows,(1)In this thesis,CNKI academic knowledge image library is used as the image material library,and an academic image data set is made as the experimental data set of this thesis.This data set refers to the production process of caltech-256 data set and combines with the classification rules of Chinese library classification number.In addition,the classification of the data set is marked by the research group by Professor Sun Wei from School of Information and Control Engineering of China University of Mining and Technology,which can ensure the accuracy of data set classification.The data set contains 20 classifications,each classification contains 50 color images of 128*128 pixels,a total of 1000 images,which can be enough for the research needs of this thesis.(2)In order to improve the accuracy and efficiency of rotating-invariant LBP algorithm in academic image retrieval,this paper improves the rotating-invariant LBP algorithm and proposes an enhanced rotating-invariant LBP algorithm.This algorithm samples the ROI before the original image is transformed into LBP pseudo-gray image.In the subsequent generation of LBP pseudo-gray map,only the pixel points in the ROI are calculated,and Harris corner points are taken as the sampling center in the ROI.The increase of this sampling process reduces the amount of calculation and data storage in the generation of LBP pseudo-grayscale map,improves the efficiency of the algorithm,and highlights the texture information of the region rich in texture information.In this paper,experiment 1 was designed to verify the optimal sampling parameters of the algorithm in this paper,experiments 2 and 3 were designed to verify the advantages of the enhanced rotation-invariant LBP algorithm compared with LBP,Uniform LBP,LBPri,CSLBP and Uniform LBPri,and experiment 4 was designed to verify the effectiveness of image retrieval method based on enhanced rotation-invariant LBP feature extraction.(3)Although the image retrieval method based on enhanced rotation invariant LBP feature extraction is very effective in the retrieval of academic image categories with rich texture information,its retrieval accuracy is low in the retrieval of image categories with rich color changes or large scales change or large angles change.To solve this problem,this thesis proposes a multi-feature fusion image retrieval method based on adaptive weight.This method integrates three feature extraction algorithms,namely enhanced rotating-invariant LBP feature extraction,quantized HSV feature extraction and BoVW feature extraction,and comprehensively extracts various underlying features of the image.Genetic algorithm is used to train the feature weights in this method,so that the feature weights can adaptively approach the optimal value in the continuous evolution,and then the retrieval process of academic images can make the best use of the three underlying features of various academic images.Experiments 5,6,7,8 and 9 were designed to verify the effectiveness of the multi-feature fusion image retrieval method based on adaptive weight.