SIFT feature reduction based on feature similarity of repeated patterns

In recognition using SIFT and SURF, matching of features extracted from both input images and learning images are involved. However, the recognition target is not always facing the same direction as that of the learning image. For this reason, recognition is performed by making it learn the image of the object you wish to recognize by capturing it from a wide variety of angles. Also, where repeated patterns are included in the object to be recognized, mismatching and misrecognition often occur because of the fact that the repeated pattern feature descriptors are similar. To avoid this issue, similarity of feature descriptors is measured using cosine similarity, and features with a specific degree of similarity are excluded. In particular, as this degree of similarity differs depending on the image, in this paper, we propose a method of seeking the optimal degree of similarity depending on the image and excluding repeated pattern features. Finally, we verify the effectiveness of the proposed method by showing some experimental results.