| The heterochromatic material detection technology could replace human to pick out heterochromatic material from the mixed material.And this technology realized the intelligent of removing heterochromatic material.It is extraordinary significance to the product quality and the competence of enterprises, and it could improve efficiency and stability of sorting enormously.The technology is widely used in industry and agriculture manufacture, military and other fields, and it is significant practicability.In this paper, detailed analysis on heterochromatic material detection is given and pattern recognition model of heterochromatic material detection is presented.The traditional image segmentation can not obtain effective features for this system. And, the algorithm needs quite a long time.Moreover,in order to achieve rapid heterochromatic location,general thought is as follows:the image is split into a number of discrimination unit, then according to the overall chromaticity characteristics of unit to determine whether it has heterochromatic material.So this problem is converted to the discriminant propertyof unit.Heterochromatic material type has a lot of randomness in the actual production.The shape,size and other features are inappropriately.And color is the simplest and most effective feature in an image.Integrated a variety of factors,features are extracted by color moment method in RGB and HSI color space. It was found that the detection effect is good when H mean, I mean and I standard deviation of discrimination unit as feature.In traditional classification, each type of samples are broadly distributed in a range. There fore, selected any sample can represent property of the class.In this study, the feature distribution of the target particles are relatively stable.However, the types of heterochromatic material can not be predicted, resulting in its features are scattered. So choosed heterochromatic material samples cannot represent the class distribution.The phenomenon is a typical One-class problem. So this paper selects a one-class support vector machine to design classifier.It assumes that all features have the same importance, but weak correlation features int the actual sample, to some extent, affect the generalization ability of classifier.So this paper presents feature weighting one-class support vector machine.The experimental results show that the detection rate of heterochromatic material by FWOCSVM has been improved, and lower false positive rate compared with the threshold and OCSVM. |
