Research On Image Retrieval Based On Linear Dimensionality Reduction And Relevance Feedback

In recent years,with the rapid development of electronic information technology and the Internet image resources,the demand for image retrieval is becoming stronger and stronger.Taking advantage of the shape,the color,the texture and the spatial information of images,Content-Based Image Retrieval(CBIR)processes and retrieves feature vectors which are formed by the low-level visual features.At the same time,CBIR conquers the problems including inaccurate description and unsatisfactory retrieval efficiency in the field of Text-Based Image Retrieval(TBIR)and has the characteristics of easy automation and intelligent.As the research on image retrieval moves along,some problems of CBIR are becoming a new hot topic.Two serious problems of “dimensionality curse” and “semantic gap” in CBIR are deeply researched in this paper.Based on the in-depth analysis of dimensionality reduction and Relevance Feedback(RF)technologies,the main work and innovative achievements in this paper are as follows:(1)In the field of dimensionality reduction technology to high-dimensional feature vectors,focusing on the issues that the Principal Component Analysis(PCA)related dimensionality reduction methods are limited to dealing with nonlinear distributed datasets and have poor robustness,we utilize the strong robustness of Angle Optimization Global Embedding(AOGE)algorithm and propose a novel linear algorithm called Robust Cosine-Euclidean Metric(RCEM)by local orthogonal projection method.Considering that Cosine metric can handle the outliers efficiently and Euclidean distance is able to maintain variance information of samples well,the RCEM algorithm utilizes the Cosine metric to describe the geometric characteristics of neighborhood and uses Euclidean distance to depict the global distribution of dataset while using parameters to balance the relationship of datasets' global and local structure.This new proposal method increases the robustness of local dimensionality reduction algorithm,helps in avoiding the problem of small sample size cases,and improves the performance and speed of image retrieval.(2)Targeting at the problem that the hidden nodes of Extreme Learning Machine(ELM)are not learning adequately,a novel classification algorithm called Discriminative Extreme Learning Machine(DELM)is proposed based on Minimum Class Variance Extreme Learning Machine(MCVELM)algorithm.The DELM algorithm introduces Maximum Margin Criterion(MMC)to ELM and analyzes both the within and between scatters of feature space in ELM hidden layer to detect a discriminative classification model – DELM,which improves the discriminant capacity of classification model.Specially,two versions of DELM,i.e.,dimensionality reduction free based version and dimensionality reduction based version are proposed.(3)Concerning the high computational complexity,low discriminant ability and insufficient image feature extraction of Support Vector Machine(SVM)related RF method,a novel DELM based RF method is proposed in this paper.This method extracts image features through color,texture and edge information of the image and overcomes the defects that image feature extraction with previous retrieval method is not sufficient.In the phase of retrieval feedback,this method utilizes DELM classification model to improve the retrieval performance of RF based image retrieval system.A large number of experiments show that,compared with other RF methods,the proposed method can significantly improve the quality of retrieval performance under the same conditions and datasets.