Configuration Recognition Of Automobile Front Face Parts Based On Improved And YOLO V3 Algorithm

The Quality Gate of the final-line station of the assembly needs to make a judgment on whether the parts on the vehicle are assembled correctly,and check whether the vehicle parts are configured correctly.At present,most workshops still perform visual inspection through operators.This kind of detection method will be affected by objective factors such as the environment changing and operators condition,which could result into low accuracy and low timeliness of vehicle inspection.Only by manual inspection,whether the parts assembled on the vehicle are correct and the quality of the produced vehicle could not be guaranteed,is reduced,and the safety of the vehicle could not be guaranteed effectively.Therefore,accurate and timely inspection of the assembly of the front end of the vehicle is good for reducing the false inspection of parts,miss inspection,workload of the inspection of operators,and improving the quality of the vehicle.In order to improve the accuracy of detection and shorten the detection time to improve the timeliness,this paper takes the final-line station of an assembly shop as an example.With the grille,fog lights and wheels rim in the front end of the vehicle as targets to check whether the parts are assembled correctly.Select SVM algorithm and YOLO V3 algorithm to identify and classify the configuration images of the front end parts of the vehicles,and improve the YOLO V3 algorithm to greatly improve the accuracy and timeliness.The main research contents are:(1)Set up an experimental platform for collecting images information of front end of vehicles,and develop a Label Image system for image labeling.In this paper,a test platform for collecting images of front end parts of vehicles is set up with the quality inspection station at the final-line of an assembly shop as an experimental point.Use this platform to collect different configuration images information of different parts of the front fend of the vehicles produced in the workshop,the parts including wheels rim,fog lights and grilles.A total of6000 part images were collected through this platform as experimental data.This paper developed the Label Image system,which can analyze and mark the collected images,and automatically generate the corresponding configuration file to prepare for the next image classification.(2)Select SVM classification algorithm and YOLO V3 algorithm to recognize the front end parts configuration of the vehicles.This paper first uses SVM algorithm to identify images.The recognition steps include pre-processing,feature extraction and classification.Pre-processing is a images operation of graying,edge detection,image enhancement,and image devoicing.Feature extraction uses the LBP algorithm to extract features from the image.When using the SVM classification algorithm to classify images,a cross-validation method is used to select the kernel function and the kernel function coefficients.Through experiments,the kernel function is selected as a Gaussian kernel function,the selected kernel function coefficient is a Gaussian kernel radius is ? = 263,and the penalty coefficient is C = 10.Secondly,the YOLO V3 algorithm was used to identify the image.The experiment was performed using the open source YOLO V3 code on the Internet.Finally,the accuracy and timeliness of the two classification algorithms are compared experimentally.Under different color vehicles and different parts,the recognition speed and accuracy of the YOLO V3 algorithm are higher than the SVM algorithm.Compared with the SVM algorithm,the YOLO V3 algorithm has hardly reduced the missed detection rate of the vehicle parts in the image when the quality inspectors are obstructing it.(3)Optimize the parameters of the YOLO V3 algorithm based on the biogeography optimization algorithm to improve the accuracy and timeliness of the algorithm.With YOLO V3 as the basic unit,the BBO optimization algorithm is used to optimize the parameter learning rate ? in the YOLO V3 algorithm,so that it can take the optimal learning rate ? every specified number of iterations,shorten the training time,and improve the timeliness of detection.Through experiments comparing the YOLO V3 algorithm before and after optimization,the recognition speed and accuracy of the BBO-YOLO V3 algorithm are higher than those of the YOLO V3 algorithm by identifying vehicles with different component configurations and different colors.For operators obstruct,experiments verify that the BBO-YOLO V3 optimized algorithm can reduce the missing detection rate caused by incomplete imagines.