CAJ | 학술논문

SIFT（Scale-Invariant Feature Transform）特征提取是通过在不同尺度DOG（Difference of Gaussian）层进行逐个像素遍历获取极值点得到，当图像分辨率较高时计算量巨大。利用DOG（高斯一阶差分）层极值点与D2（高斯二阶差分）层的近零点之间的对应关系，提出基于近零点模板的SIFT特征提取算法，其中近零点判定阈值利用图像熵动态获取。由于模板的限制致使SIFT特征提取的范围缩小，极大降低了计算及时间复杂度。实验结果表明，相对于经典SIFT算法，提出的算法不仅保持了其较高的鲁棒性，而且大幅提高了特征提取效率。
SIFT（Scale-Invariant Feature Transform）특정제취시통과재불동척도DOG（Difference of Gaussian）층진행축개상소편력획취겁치점득도，당도상분변솔교고시계산량거대。이용DOG（고사일계차분）층겁치점여D2（고사이계차분）층적근영점지간적대응관계，제출기우근영점모판적SIFT특정제취산법，기중근영점판정역치이용도상적동태획취。유우모판적한제치사SIFT특정제취적범위축소，겁대강저료계산급시간복잡도。실험결과표명，상대우경전SIFT산법，제출적산법불부보지료기교고적로봉성，이차대폭제고료특정제취효솔。
Feature extraction of scale-invariant feature transform (SIFT ) is implemented by obtaining the extreme points through one-by-one pixel traversal in different scales layers of difference of Gaussian (DOG).As the image resolution increases,the calculation becomes enormous.By employing the corresponding relation between the extreme points in DOG layer and the nearly zero points in Gaussian second-order differential (D2)layer,we put forward the SIFT feature extraction algorithm which is based on nearly zero template.In it the nearly zero point discriminant threshold is obtained by making use of the image entropy dynamics.The limitation of the template greatly reduces the scope of SIFT feature extraction,calculation and time complexity.Experiment result shows that in contrast with the classic SIFT, the algorithm proposed in this paper preserves its higher robustness and dramatically improves the feature extraction efficiency as well.