一种基于SIFT的仿射不变特征提取新方法

摘要

The extraction of image local features is a key problem in the field of image understanding and computer vision. SIFT has been widely used owing to its distinctiveness and robustness. However, SIFT is not robust to affine deformations, because it is based on DOC detector which extracts circle regions for keypoint location. Besides, SIFT descriptor is represented by a 128 element feature vector, and when it is used for image registration especially for the case that there are many keypoints in the image, the matching speed will be slow and storage requirement will be huge. In order to overcome these disadvantages, a novel descriptor named HA-DR-SIFT ( Hessian Affine-Dimensionality Reduction-SIFT) that is robust to affine deformations is proposed. Firstly, Hessian-Affine Detector inBtead of DOG detector for keypoint location detection is used to make the elliptical region affine invariant. Then, the PCA or NLPCA method for new descriptor's dimension reduction is used to improve successive computing efficiency. The experiments show that the novel descriptor is not only invariant to affine changes, but also has higher matching speed and lower storage requirement than SIFT.%图像局部特征提取是图像理解及机器视觉领域一个非常关键的问题,其中SIFT特征因具有良好的显著性和鲁棒性而得到广泛应用.但是,SIFT采用DOG检测子,定位的特征区域为各向同尺度变化的圆形区域,故其只具有尺度不变性,并不具备仿射不变性.此外,SIFT采用128维特征向量表示,当在图像特征点较多情况下进行匹配实验时,存在存储空间大、匹配耗时多等缺点.针对这两个问题,本文提出一种新的仿射不变特征提取方法,即HA-DR-SIFT(Hessian AffineDimensionality Reduction-SIFT).首先,用Hessian-Affine检测子代替DOG检测子,使提取的椭圆图像区域满足仿射不变性需求;其次,用PCA或NLPCA方法对128维特征向量进行降维处理,提高后续运算效率.实验表明,新方法不仅具有良好的仿射不变性,而且在匹配时间和存储空间上优于SIFT算子.