目的 作为计算机视觉的热门研究方向,局部不变特征算法的发展已趋于成熟、稳定,然而目前几乎所有特征点提取算法都没有给出特征点的精度指标.针对这一缺陷,提出一种特征点精度指标-特征点波动区间.方法 性质稳定的点在干扰条件下仍具有较好的精度,即小范围的波动区间,因此,以当前最热门的SIFT(scale-invariant feature transform)特征点为例,在图像加入噪声,发生光照变换,发生模糊变换以及同时进行噪声、光照及模糊处理这四种情况下分别分析同一算法提取的不同特征点的波动情况,进而得到不同特征点的波动区间.结果 实验得到16个稳定检出特征点,其中点2,3,4,11,13这5个点可以在不同干扰条件下的波动范围都较小,而点2则只在模糊条件下波动较小,在其余干扰下波动较大.特征点虽然已经过特征提取,但仍具有不同的波动区间,其优劣也不尽相同.不同的特征点的波动区间并不相同,但仍有一部分特征点在不同干扰条件下均保持较高的提取精度.结论 波动区间能很好地表征每个特征点的提取精度.由于此前只有针对特征点算法的评价指标,而没有针对特征点自身性质的评价方法,因此本文提出的特征点波动区间将为诸如设备标定、视觉测量、精简特征库等相关后续工作打下良好基础.

Objective As a popular research direction in computer vision, the development of local invariant feature algorithms has become more and more mature and stable. But now, almost all the feature point extraction algorithms cannot give the accuracy index of feature points. In fact,the precision of feature points'position is requested in many areas,such as device calibration and visual inspection. To solve this problem, this paper proposed a feature point precision index-feature point's range. Since there are always disturbances in the process of acquiring images,it is difficult to obtain absolutely perfect images and feature points. This paper defines the feature point's range as the range of a feature point's fluctuation in different conditions which is stable in different disturbances. Method This paper choses the most popular algorithms of local invariant feature named SIFT(Scale-invariant feature transform)as an example. To make the result more intuitive and clear,the experiment selects those pictures whose backgrounds are simple and objectives are clear. This paper analyses the fluctuation of feature points in noise conditions,fuzzy conditions ,light transform conditions and all those disturbances which are the common interferences in actual operations to achieve the fluctuation range of different feature points. First,this paper establishes the experimental galleries. We make not only a noise gallery, light treatment gallery and fuzzy diagram gallery but also a gallery which contains the three disturbances. It is worth noting that considering the randomness of noise,we generate 100 pictures for one noise intensity. Secondly, the stable feature points which can be detected in all disturbances are found by matching. We get 16 points in this experiment which has its own point cloud. Once more, because different points fluctuate differently in different situations, we use a circle to fit every point's fluctuation in different conditions. It means we use a circle to fit every point's cloud. The radius of those circles can characterize those feature points' ranges. Last but not least,this paper uses histograms which is very intuitive to describe every point's fluctuation. In addition,those points' coordinates are provided. Result In this experiment we gain 16 stable points. This experiment shows that the fluctuation ranges of different feature points are different,but there are still a part of feature points whose precisions are higher. The points whose fluctuations are smaller in the case of the presence of interference can be considered as better and more accurate points. Selection of feature points is based on the following work. If the requirement of precision is higher,a lower threshold should be designated. Conclusion Therefore,fluctuation range can characterize the precision of different points very well. This can underpin the related work. Although this paper choses SIFT feature point as an example,other local invariant feature points' have similar properties. This paper provides an idea and method to study feature points' properties to be helpful to the related work.