Research And Development Of Industrial Grade PVC And Its Identification Acceleration Mechanism

PVC leather is an industrial product that is widely used in daily life.In the production process of industrial PVC leather,due to environmental and equipment factors,the leather in the production line will inevitably have defects.Therefore,it is necessary to check the quality of the finished PVC leather products.Inspection to ensure that defective leather does not enter the market.At present,the quality inspection link of domestic PVC leather production line mainly adopts manual inspection of leather surface defects to judge whether the quality is qualified.On the one hand,this inspection method has the influence of human subjective factors and is inefficient.The method cannot detect and stop the source of the defect in time and cause losses.In view of the more mature development of machine vision technology and target detection technology,this thesis designs an industrial PVC defect detection system based on machine vision and deep learning target detection technology,aiming to apply target detection technology to industrial PVC production lines for in-production PVC inspection Leather surface defects,quality inspections are carried out at the same time as production to avoid production waste caused by defects.Due to the fast speed of the leather production line,this thesis uses the YOLOv5 algorithm,which is currently excellent in detection speed,to detect defective targets,and studies the acceleration mechanism of the algorithm to identify targets,so that the detection speed can adapt to the production line speed.The main work of this thesis is as follows:1.According to the field investigation of the PVC leather production line,summarize the environment and requirements of the production line,analyze the structure and functions of the system in detail,and divide the system functions into five modules,namely online detection module,operation statistics module,and product management.module,camera management module and system configuration module.Based on the Windows platform,this thesis completes the UI development of each module of the system combined with the Duilib interface library technology,uses SQLite to add,delete,modify and check system data,uses the Pluma framework to realize the plug-in of the function module,and completes the camera control plug-in based on this plug-in technology.And the development of the defect target detection plug-in,for the camera control plug-in technology,this thesis takes the control of DALSA camera as an example,and expounds the basic process and implementation method of controlling industrial cameras.For the defect target detection plug-in technology,this thesis combines the OnnxRuntime framework to realize model inference.2.Research the inference acceleration mechanism of deep learning,and lighten the model based on the YOLOv5 algorithm.In this thesis,combined with the advantages of the lightweight network MobileNetV3,the MobileNetV3 network replaces the network backbone of the YOLOv5 algorithm.In order to analyze the feasibility of replacing the network to speed up the detection of defect targets,this thesis uses the PVC leather surface defect pictures obtained by on-site investigation as the data set,that is,8700 PVC leather defect pictures obtained after image preprocessing.After the network,the detection accuracy of industrial PVC leather defects has decreased,but the detection speed has been significantly improved.In combination with industrial production requirements,the increase in detection speed in exchange for accuracy is within an acceptable range in the industry,reaching the speed of detection of PVC leather defects.the goal of.This thesis implements a hardware acceleration scheme for model inference based on the CUDA framework and TensorRT framework,and conducts comparative experiments to obtain the feasibility of using GPU acceleration.Through the implementation feedback of the industrial site,the industrial PVC defect detection system developed in this thesis improves the detection efficiency of the defect target in terms of leather quality detection,and is suitable for real-time quality detection tasks in the production process of industrial PVC leather.