一种目标识别中基于关键点的仿射不变矩

针对常用仿射不变矩对轮廓边缘信息敏感的特点,提出了一种基于关键点的局部仿射不变矩,以分割出来的物体灰度图像为基础,首先计算出物体的质心,然后以质心为扩展点向周围引伸出多条射线,寻找每条射线方向上的最近灰度极值点,将所有灰度极值点当作关键点集合,并按照文中提出的计算仿射不变矩的方法提取出多阶不变矩,将其当作神经网络的输入向量,放入已经训练好的神经网络来达到识别物体的目的,将此方法应用在了飞机图像的识别上,实验结果证明此方法在物体轮廓分割不完整和有噪点污染的情况下能保持很好的稳健性,简单有效并具有较广的应用范围.

Affine invariant based on determinant points in object recognition

The common affine invariants were very sensitive to the edge of the image, so a new method entitled "an affine variant based on the determinant points in image recognition" was proposed. First the centroid of the separated object image was computed, and then many line segments were derived through the centroid, at the end the nearest extreme grayscale points were found on every line. In order to compute the affine invariants, the extreme points were used to build up a collection of determinant points. The affine invariants of the collection could be served as an input vector of the trained neural networks to distinguish whether the source image was the destination image. The method was applied to the plane recognition and was proved to keep highly stable even if the object contour was ill-segmented or noisy. It is much easier and effective compared with the traditional methods and has a very wide scope of application.