Research And Implementation On Defect Detection System Of Press Pump Based On Machine Vision

The press pump is an important lotion container bottle in people's lives.In the production of enterprises,it is inevitable that due to the backward production technology and the aging of the injection molding template,the surface of the press pump will have defects such as lack of glue,oil stains,and scratches.In order to improve the competitiveness of enterprises,this paper designs a fully automatic and high-precision press pump defect detection system based on machine vision technology to complete the qualitative classification of products,identify the type of defects and obtain their location information.First of all,according to the design goal and detection goal of the system,a set of mechanical transmission type hardware detection platform is designed.The main body structure of the platform includes S-shaped discharge chute,multi-station rotary chuck,up and down telescopic rail type pulley fixture,polygon mirror reflection imaging detection station and sorting station.Secondly,this paper introduces the visual inspection program design and software function development process.Visual inspection is the core module of the system.The selection of hardware devices such as light sources,industrial cameras,optical lenses,etc.is completed by calculating relevant parameters,and the appropriate lighting and industrial camera layout solutions are designed according to the image imaging quality.In the system software part,the software function modules are developed from the application analysis layer,the middle platform layer,and the hardware processing layer,including pages such as image display,qualitative classification,defect detection,data analysis,parameter setting,and human-computer interaction.The software and hardware platform realizes the processes of image acquisition,detection,sorting,and information storage through the transmission protocol.Finally,according to different detection tasks,the paper studies two types of defect detection algorithms in detail.First,the classification task of the push pump is completed at the image level through traditional image processing techniques.Firstly,the imagepre-processing methods are studied by comparison.According to the distribution characteristics of the defects,extract the images of the pump top,pump head,body,and tail pipe respectively,and design corresponding detection algorithms to classify the images.Including pump top squeeze deformation and oil scratch detection,pump head screw teeth detection,body oil scratch detection,tail pipe upside down detection.Second,the deep learning target detection algorithm Faster RCNN is used to complete the defect identification and positioning tasks of the press pump.Combined with the research task and the shape characteristics of the defects,the basic network was selected as Res Net101's Faster RCNN detection algorithm.Secondly,the algorithm's feature extraction network,candidate region generation network,and ROI detection network are studied.In order to improve the performance of the algorithm,this paper corrects the network structure of the algorithm,including combining the FPN structure to improve the feature extraction ability,using deformable convolution to improve the defect learning ability,and the pre-selection box screening mechanism based on Soft NMS and Box Voting to improve the defect positioning accuracy.After training the model,the detection results before and after the algorithm correction are compared.Experiments show that after the algorithm is corrected,the detection effect of small targets or irregular defects is better,and the detection rate of different types of defects is obviously improved.The test results of the inspection system show that the qualitative classification algorithm and the defect detection algorithm meet the system design requirements in each inspection index.The detection system studied in this paper has a good reference value for industrial applications.