Research On ELM Based Tri-training For Relevance Feedback In CBIR System

With the rapid development of data acquisition technology, storage and multimediatechnology, many applications have generated a lot of image data. How to quickly andaccurately identify the image from the various types of ultra-large-scale image database is thehot topics of the current pattern recognition, information processing and related fields.Content-based image retrieval refers to the image database retrieval based on the underlyingfeatures of the image, retrieve images that match the query image. In the field of imageretrieval, the most important problem is the "semantic gap", that there are significantdifferences between underlying features of the image and human’s understanding of thecontent to the image.To introduce the relevance feedback mechanism in image retrieval is one of the effectiveways to solve the problem of "semantic gap". Due to the higher dimension of image features,the traditional retrieval algorithm computing speed is slow, and therefore to improve theaccuracy and improve the speed of the algorithm is very important. In order to solve theseproblems, in this paper we propose a semi-supervised feed forward neural network withrelevance feedback algorithm, and further, we study the ranking problem of the search result.The main contents of this dissertation are as follows:(1) This thesis presents a fast classification algorithm based on Tri-training ELM forimage database classification, in order to solve the current CBIR system to be running slow.(2) Based on Tri-training ELM classification algorithms, we propose a new relevancefeedback strategy to get more satisfactory search results.(3) In this paper, we have improved the similarity integral reordering algorithm, andcombined the improved with relevance feedback, we obtain a better sort results.The proposed methods were tested on the Corel image database, the results show that ouralgorithms are effective.