Supervised Learning Based Image Local Interest Points Detector

In recent years, the research of most image local interest points methods focus ondesigning and learning an e?ective local interest point descriptor algorithm to charac-terize a local interest point, and few research focuses on the algorithm of local inter-est point detector. However, how to identify stable local interest points is still a verychallenging problem. Traditional local interest point detector algorithm is based onhand-craft strategies to select stable local interest point, in which it is hard to choosea common strategy for all the images. In this master thesis, we adopt a supervisedlearning based data-driven approach to improve the algorithm of selecting stable localinterest points. We mainly focus on the algorithm of Scale Invariant Feature Transfor-m (SIFT). We propose a new algorithm which is called Rank-SIFT. In this algorithm,we adopt learning to rank framework, design a set of di?erential, grads and edge re-sponse features from each point, and de?ne the stability score of a point across imagescontaining the same visual objects. We train Rank-SVM model on extracted trainingdata based on the proposed features and stability score. This model is used to predic-t the stability of each point, which is helpful to select stable interest points. At thesame time, we construct images database for evaluation, and de?ne the standards ofRepeatability and Matching score measures. And also, we carry on image retrievaland category classi?cation experiments on the Oxford and PASCAL database, respec-tively. The experimental results show that compared with traditional SIFT method,Rank-SIFT can substantially improve the stability of detected local interest points. Wealso analyze the importance of each designed feature in the stability score predictionusing sparse coding algorithm in the experiments. Finally, an online web image localinterest point detection system is implemented, and also a open source software for Rank-SIFT algorithm is released.