Performance Evaluation for Scene Matching Algorithms by Svm

Scene matching is the process of locating a region of an image with the corresponding region of another image where both image regions represent the same scene. Although a lot of algorithms have appeared on scene matching, performance analysis is usually based on simple statistic experiment and performed simply and visually, and little attention has been given to evaluate performance of different algorithms. In order to choose suitable algorithms and improve the performances of the algorithms, we present a novel performance evaluation method for scene matching algorithms based on support vector machine (SVM), which can partly show interact-effect of numerous similarity measure factors and find a dependency link between two correlative images. The method is described with a three-step procedure. Firstly we build samples data set using similarity measure descriptors of image pairs. Then decision function is obtained through training and testing process with input of samples data. Finally, we adopt result of SVM classification to evaluate two classical algorithms: normalized cross-correlation algorithm and Canny-based edge extraction algorithm. The experimental results show that this method holds the capability of automatic decision ability for performance evaluation and high ratio of correct prediction. * Corresponding author