Study On Image Retrieval With Relevance Feedback Based On Improved Random Walks

As we all know, there is a semantic gap between what he perceives as similar and whatthe provided low-level similarities classify as similar. The set of visual features the metricused to model the user’s perception of image similarity occupy a very important position inimage retrieval, and the efficiency of content-based image retrieval (CBIR) system is largelydepend on the set of visual features and the metric used to model the user’s perception ofimage similarity. In order to explore the intention of the user and improve the system’sperformance, the feedback mechanism is incorporated into the image retrieval system.Relevance feedback is an interactive retrieval method, put the subjectivity of the humanperception of the images into image retrieval and is the learning process. Although it is arelatively complete feedback retrieval system, the importance of each dimension characteristicand spatial relations between images are not considered. To solve the above problem, thispaper proposes an improved algorithm. Extensive experiments on different real datasets withseveral image features show the superiority of our method over different recent approaches.This paper discusses the following two issues:Firstly, feature weighting using Laplacian Score, which has the ability to maintainstrongly a local structure. Fully reflect the importance of each characteristic dimension.Secondly, k-nn connecting to Random walk, k-nn classify the training samples in thefeature space, samples of the same class have a high degree of similarity. By calculating thesimilarity of the unknown sample with a known class category between samples. The basicidea is to find k nearest neighbors of the test sample in the training s ample, to determine theclass of the test sample according to this class of k nearest neighbors.Combine this aspects with image retrieval method based on random walk, not onlyimprove the importance of distinguishing characteristics of each dimension, but also makefull use of the spatial relationships between images. The algorithm explores the intention ofthe user and improves the efficiency of the CBIR. Therefore, this paper provides a newexploration for content-based image retrieval.