Image Retrieval Uses SVM-Based Relevant Feedback for Imbalance and Small Training Set

Abstract

There are many image retrieval systems that use the SVM-based relevance feedback approach to reduce the gap between low-level visual features and high-level semantic concepts. However, the performance of these systems is low due to the lack of two issues: first, the imbalance of the training set. Second, the size of the training set is very small compared to the dimension of the feature. In this paper, we propose the image retrieval method, SVMITS (SVM-based relevant feedback for class imbalance training set), to overcome the above limitations. Our proposed method solves the first problem through resampling techniques and the second is through reducing the number of dimensions of the features. Finally, we incorporate resampling techniques, dimension reduction, and ensemble-based learning techniques to improve the accuracy of image retrieval using SVM-based relevant feedback. To illustrate the effectiveness of our proposed method, we provide empirical results on a database of 10800 images.