Target Recognition And Location Based On Machine Vision And CNN

Target recognition and positioning,as an important branch of the visual field,has an indispensable position in industrial production and is an important measure of industrial automation.The paper is based on machine vision and the emerging CNN(Convolutional Neural Network).The purpose is to identify the target category information and the twodimensional position information of the target object through the combination of calibration and deep learning with the original vision system.Information is transformed into threedimensional information to achieve target recognition and location in industrial scenarios.Firstly,the paper studies the camera calibration algorithm.Due to the advantages of high detection accuracy and stable performance,the checkerboard is selected as the target to achieve camera calibration.A special target recognition algorithm is designed.Two stable binocular vision calibration methods are used.The camera performs a single target setting.Finally,the feature points of the left and right camera images are matched by the principle of polar line constraint in the feature point matching process to realize the calibration of the binocular camera.Then,for the target detection algorithm,the basic theory needed for the convolutional neural network is expounded.Because SSD(Single Shot MultiBox Detector)is different from other target detection algorithms,it can detect target objects relatively quickly and ensure that it stands out in the target detection field while ensuring high accuracy.However,the harsh hardware requirements of the SSD network structure and the length of the pre-training time make some small real-time systems prohibitive.To this end,we deeply analyzed the basic network VGG16 and the SSD detector different from VGG16.This paper retains the SSD detector,replaces the traditional convolution with deep separation convolution,and introduces the BN and residual network structure to modify the VGG16 network structure,which deepens the original network depth and improves the generalization ability of the network structure.The modified network and the original SSD are trained and tested on the PASCAL VOC data,which proves that the modified network accelerates the training and testing speed of the network and reduces the dependence on the hardware under the premise of ensuring accuracy.Finally,the modified SSD network is used to train its own data set,and the threedimensional information of the target object is obtained by combining the two targets,and a better effect is obtained.In summary,for the target recognition and location in industrial scenarios,the algorithm mentioned in this paper has obvious effect improvement,and has certain research significance and late application prospects.