Research On Multi Object Recognition In Micro-Vision System Based On Neural Network

With the development of robot technology, the operation objects extend to micro field.Now, the technology of microassembly robot gradually comes to maturity. Depend onvisual servo, microassembly robot can get the image coordinates and space coordinates ofthe operation objects, which is the precondition of the automatic assembly manipulator.Research on multi-target recognition based on neural network under micro vision presentedin this paper is an important component part of the microscopic vision servo control system.And recognizing the objects exactly is the basic of microscopic vision servo controlsystem.This paper based on target assembly robot as a platform, and realize the microscopicvisual multi-target recognition, which including the pretreatment of the multiple targets,target segmentation, feature extraction and target recognition. Firstly, in this paper weintroduce the system structure of target assembly robot. Based on the characteristics ofmicroscopic visual image, a serious of image preprocessing was done, which includingimage binarization, average filtering, image enhancement, edge detection. After severalcommon edge detection method of comparison, we use canny edge detection method to getobject edge, and use the clustering method to conduct target segmentation. According tothe structure characteristics of microscopic visual targets, the invariant moment methodwas used to conduct characteristic feature extraction. Formal methods like invariantmoment method can’t recognize and get the characteristic of the target which was partlycovered. According to the problem in Micro-Vision System, a new method of targetrecognition in Micro-Vision system which based on Hough transform and templatematching was proposed in this paper.According to the advantage of neural network in classification, we use the neuralnetwork to conduct target recognition. The input of the neural network classifier was the characteristics quantity of each target. Then, use the samples of each target to train theclassifier. After training, the system can realize multi-target recognition and classification.In addition, the system structure of multi-target recognition system based on neuralnetwork was designed. The experiments proved that the classification of the methodmentioned previously was valid.