Tensor based feature detection for recognition of poorly illuminated objects
Fernando Merchán, Filadelfio Caballero, Damien Rousseau, Héctor Poveda · 2014
In this work we present an extension of the SIFT algorithm to color images. In the extrema detection stage, an energy level descriptor based on the color tensor of the image is computed and used to locate keypoints candidates. Then, in the description stage, the color gradient magnitude and orientation of the samples around the keypoint are used to compute an orientation histogram to create the keypoint descriptor. A comparative study is carried out between the proposed algorithm and the classic SIFT and the C-SIFT algorithms in several illumination settings. The proposed algorithm presents a better performance in terms of accuracy when objects are poorly illuminated.