Keypoints-based image matching algorithms have proven very successful in recent years. However, their execution time makes them unsuitable for online applications. Indeed, identifying similar keypoints requires comparing a large number of high dimensional descriptor vectors. Previous work has shown that matching could be still accurately performed when only considering a few highly significant keypoints. In this paper, we investigate reducing the number of generated SURF features to speed up image matching while maintaining the matching recall at a high level. We propose a machine learning approach that uses a binary classifier to identify keypoints that are useful for the matching process. Furthermore, we compare the proposed approach to another method for keypoint pruning based on saliency maps. The two approaches are evaluated using ground truth datasets. The evaluation shows that the proposed classification-based approach outperforms the adversary in terms of the trade-off between the matching recall and the percentage of reduced keypoints. Additionally, the evaluation demonstrates the ability of the proposed approach of effectively reducing the matching runtime.