【摘要】 为识别任意角度扭转的货品商标图像,提出了对扭转鲁棒的SIFT图像局部匹配算法。通过模拟商标图像的垂直和水平多角度扭转,对数据库中所有商标图像提取多组SIFT特征。提取查询图像的SIFT特征,并与库中特征进行匹配从而识别商标。提出了一种仅采样少量角度来近似模拟任意角度图像扭转的方法。实验结果表明,该方法在1000幅商品图像集上正确匹配率高于传统的SIFT算法,可从各种角度的图像中有效地检测到商标图像。更多还原

【Abstract】 To match product logos with various viewpoint changes in realistic images,a rotation-invariant SIFT matching algorithm is presented.Firstly,logo images in the database are rotated to several simulate angles in both horizontal and vertical directions.And multiple sets of SIFT features are extracted from those rotated samples.Then,the SIFT features are used to match features of query logo image and database to recognize the type of query logo.Approach of simulation of several rotation changes to achieve full rotation-in-variance is proposed.Experiments on a dataset of 1000 product images indicate that the proposed method achieves higher precision than traditional SIFT method,which demonstrates that proposed method is applicable for logo detection applications.更多还原