High Speed Real-Time Foreign Recognition Based on Hypersphere One-Class Support

Aim at the foreign randomicity and the color difference between the regular material and the foreign in high speed real-time foreign recognition, an approach to real-time foreign recognition based on hypersphere one-class support vector machines is presented. In this method, regular material are considered as training samples, and OC-SVM is used to obtain the color distributing of the regular material, that is to find out the foreign whose color are not in the range of the regular material. To train and simplify OC-SVM, centrifugal coefficient omega is presented, and Zoutendijk fastest decline method is adopted to search for working set. The result of experiment shows, this method has good performance in foreign recognition, especially for some foreign whose color is similar to that of regular material. Contrast to Gaussian distribution fitting method, the general accuracy of recognition is 8~10% higher.