基于非监督特征学习的兴趣点检测算法

由于兴趣点是图像中的基础、关键特征,因此兴趣点检测是图像配准、图像检索以及图像识别的关键步骤。基于兴趣点对于图像特征响应较为强烈的特性,结合非监督特征学习算法可以自主地从无标签的样本中提取特征的思想,提出了UFL-ID兴趣点检测算法。该算法无监督学习了图像的底层特征,对特征进行信息量和各向同性的评价,并利用特征的卷积响应及评价参数寻找图像中的兴趣点。与其他常见的兴趣点检测算法的对比实验表明,该算法具有良好的重复性与抗噪能力。

Unsupervised Feature Learning Based Interest Point Detection Algorithm

Interest point is of great importance in digital image processing as a kind of critical feature at low level.So the interest point detection is the committed step in image registration,image retrieval and image recognition.In this paper,an unsupervised feature learning based interest point detection (UFL-ID) was presented based on the fact that interest points have stronger feature convolution response than others.The new UFL-based interest point detection algorithm firstly learns low level features in digital images,evaluates the information content and isotropy of learned features,and finally uses features and its evaluation to find interest points.The comparison result demonstrates that using UFL produces great improvements of repeatability and anti-noise property.