Research And Development Of Common Appearance Defects Detection System For Injection Molding Parts

Manual detection is the main way to detect appearance defects of injection molding parts in most enterprises.These methods are inefficient,unstable,and have poor flexibility and adaptability.The machine vision based method is the future of defects detection.Therefore,the machine vision based defects detection systems and detection methods were studied.(1)Following the actual detecting processes,an appearance defects detection system for injection molding parts was built.The framework of the system was designed,and the optical equipments were well picked.Then,following the idea of modular designing,the architecture of the supporting software for the detection system was designed,and the key features of each module were developed with.Net platform,which achieved the automatic detecting.(2)To suppress the image background that had great influence on detecting defects,an image preprocessing flow based on image enhancement,image reconstruction and image binarization was studied.Taking the scratches and bright spots as the objects,the methods with appropriate parameters were determined.With the proposed image preprocessing flow,the purpose of suppressing background texture and extracting defect regions was achieved.(3)To overcome the faults such as strong dependence and poor flexibility of artificial features based defects detection method,the convolution neural network was applied.The network structure was designed and the structure parameters,training parameters and training dataset size were determined experimentally,making an ideal defects classifier.The classifier combined with the image preprocessing flow above made up the defects detection method.(4)The proposed defects detection method was verified by experiments.Firstly,the positive effects of image preprocessing and the superiority and reliability of convolution neural network were showed by experiment results.And the correctness and flexibility of the proposed method under multi-classification task were proved by experiments.Finally,the detection results of phone shell parts showed the correct rate of the proposed method was above 95%,which implied a preferable performance and made it a good application prospect.