Research On Image Retrieval Technology Based On Multi-feature And Relevance Feedback

With the rapid development of the Internet and modern electronic equipment, thenumber of images in the internet has reached a massive scale. It makes thecontent-based image retrieval become one of the research focuses. The traditionalmethod has the semantic gap problem between low-level visual features and high-levelsemantic features. How to effectively solve the semantic gap problem has become anurgent problem.Content-based image retrieval is an interdisciplinary, involving image processing,pattern recognition, machine learning and so on. This thesis proposes an image retrievalmethod by improving feature extraction and using the word model. In order to solve thesemantic gap problem, this method introduces the relevance feedback technology. Theresults mainly include:1. The traditional SIFT (Scale Invariance Feature Transform, SIFT) feature has aproblem of high dimension, leading to a high degree of computational complexity. Anadaptive threshold SIFT feature extraction algorithm is proposed. In order to detectedthe noise-sensitive low-contrast points, the threshold is set according to the imagecontents, and this can effectively control the number of SIFT feature. By this process,the computational complexity can be reduced, and the efficiency of image featureextraction and matching can be improved.2. The word model is widely used in text retrieval field. Considering the betterretrieval result of the work model, this thesis introduces it to the image retrieval. Firstlyextracting the SIFT features from images, clustering them by the AP (AffinityPropagation, AP) clustering algorithm. The centers of the clusters are visual words.Secondly based the visual words the images are represented as the visual wordhistogram. Finally image retrieval is finished by calculating the similarity between thehistograms. The experiments show that the visual word model based on AP clusteringalgorithm has a better result in image retrieval, and it presents a good stability indifferent kind of image.3. In response to the semantic gap in the image retrieval, this thesis proposes a relevance feedback technology based on SVM (Support Vector Machine, SVM).Considering the feedback of the users as the training sample, the system adjusts theparameters by machine learning theory to enhance the accuracy. Through researchingthe SVM theory, this thesis proposes the SVM-based relevance feedback. Theexperiments show that it can significantly improve the accuracy of image retrieval, andsolve the semantic gap problem to a certain extent.