A New Approach for Image Classification Applying Reduction of Colored Keypoints

There are algorithms for feature extraction such as SIFT and Opponent-SIFT, which detect and describe keypoints. In image classification, it is common to have an image dataset. Therefore, when using an algorithm to detect and describe local features in a set of images, the number of keypoints detected by class can be disproportionate. This paper presents a novel approach to reduce the number of keypoints (and colored keypoints) in images after the feature extraction process so that computer vision techniques can be applied to image classification problems. This approach uses Zipf's Law and the Pareto Principle to conduct the new strategy to reduce keypoints. An experiment was conducted comparing four different strategies. Results are encouraging, and the proposal opens new paths for keypoints reduction and syntactical pattern recognition. The classification reached an F-Measure of 76,8%, and the computer performance (execution time) has increased from 9 to 1900 times.