Corner Detection by Neural Networks

• Main Authors: Kuo-Fun Tzeng, 曾國芳
• Other Authors: Du-Ming Tsai
• Format: Others
• Language: zh-TW
• Online Access: http://ndltd.ncl.edu.tw/handle/59090606120381780292

碩士 === 元智大學 === 工業工程研究所 === 83 === A new corner detection method based on artificial neural network is proposed.In this research,two neural network models are considered. One neural model participates in detecting corner points of objects, and the other in the detecting of tangent points and inflection points of objects.The input features of the first neural model are the coordinate vector of the forward segment or the backward segment with respect to a given boundary point. The output features of the first neural model is the angle between the X- axis and the forward segment (or backward segment). We use the difference of the angle between the forward segment and the backward segment as the curvature of the boundary point. If the curvature of a point is smaller than a given threshold and is a local minimum, this point is identified as a corner point. Tagent and inflection points are typical feature points of most man-made industrial parts. Because the curvatures of the tangent and inflection point are not explicit, we can not directly use the magnitude of the curvature to detect the tangent and inflection points. The chage of the curvature in the neighborhood of a given boundary point of these two feature points is different from that of non- tangent and non- inflection points. Therefore, we use the sign pattern of the curvatures as the input features of a second neural model. The second neural model only responds to tangent and inflection points. The effectiveness of the detectors has been demonstrated by experimental results of various laboratory scenes.