Key Technologies Research On Intelligent High-speed On-line Foreign Material Recognition And Sorting

Foreign recognition and sorting system is very important in modern logistics and product line, this system could replace human to pick out the foreign bodies from the mixing material. It is extraordinary significance to the product quality and the competence of the enterprises, and it could improve efficiency and stability of the foreign sorting enormously. The technology of high-speed on-line foreign recognition is widely used in industry and agriculture manufacture, madicine, food safety and other fields, and it is significant practicability.Aim at the difficulty of high-speed on-line foreign recognition, and on the research platform of tobacco on-line foreign eliminate system, some key technologies of high-speed foreign recognition and sorting are studied intensively to improve the foreign eliminate rate and reduce the mis-recognition rate. Firstly we made some experimental research in different detecting means, and designed the high-speed on-line foreign recognition system based on machine vision. The basic theory of pattern recognition was studied, and pattern recognition model of high-speed on-line foreign recognition was presented. Some foreign is hard to recognition because of its color similar to regular material and the background, so the unit recognition method was provided to solve this problem. Aim at limited samples learning and hard sampling of foreign body because of its complexity and variety, Hypersphere One-class Support Vector Machines was presented, and the D-QDPSO was presented to solve the optimization of OC-SVM. The mis-recognition rate was reduced by using OC-SVM. We studied the intelligent on-line learning technique in high-speed foreign recognition, and made the intelligent on-line learning model of foreign recognition, and presented the mnemonic self-learning and Self adapting dynamic threshold method. After studied the characteristics of liner CCD images, we presented the mechanism of on-line self-learning and self- adapting debug to liner-CCD camera. In the end, we applied the high-speed on-line foreign recognition technologies to tobacco on-line foreign eliminate system, the practices shows, the system meet the design requirements and the demand of roboticized production.The primary innovations of this paper are as follow: ①Aim at the ploblem that the method based on pixel recognition can not distinguish some foreign body whose color is similar to the regular material or background, the holistic unit recognition method was presented. In this method, the gray scale uniformity of unit was calculated to obtain a new unit character, smooth foreign can be recognised well by using it. To some foreign whose texture is evident, we use FFT to analyze its texture characters. Aother unit recognition method is self-adapting clustering analysis method, in this method, fewness colors can be the token of this unit.②A Hypersphere One-class Support Vector Machines method using in foreign recognition is presented. In this method, centrifugal coefficientωis provided, asωis a fixed value and could be calculated by SVs, the fussy process to calculate is avoided. Zoutendijk fastest decline method is adopted to seek the working set in the improved SMO, the most convergent couple of samples would be the subclass to be optimized. The datas of experiment inducate that the computation time has reduced by 20-30% as compared with that of original SMO. The general accuracy of recognition is 8%-10% higher than dimension normal distribution fitting algorithm, especially in the condition that learing samples are limited.③Joint with PSO, D-QDPSO was provided to solve OC-SVM. In the initialization, the position of one particle could be nearer to the global optimum solution; and the boundary points of subjected plane were concerned as the initialized position of other particles, so as to make the searching area wider. The new position of the directional particle is calculated based on the current global best point(gBest), which optimized direction conforms to Zoutendijk fastest decline method principle. The experiment shows, the operating speed is 2 times faster than that of standard PSO; relatively, the number of SVs is less than that of SMO, and it would make the generalization better and mis-recognition lower.④The mnemonic self-learning and self adapting dynamic threshold method were presented. By counting the sampling odds during a period of time, the material would be judged to be foreign or regular material; By counting the statistic of holistic average color value during a period of time, the threshold is real-time adjusted to fit the change of the light. By studied the characteristics of liner CCD images, we presented the mechanism of on-line self-learning and self-adapting debug to liner-CCD camera. In this method, x-coordinate is added to learning samples as a characteristic, by checking the color variety of one x-coordinate during a period of time, estimate that the imaging and recognition are affect by dust or not, so that the foreign body would not be taken as the regular material to be learning, and the recognition rate is improved.