A Method For Rapid Detection Of Surface Defects In Stamped Parts Based On Residual Neural Networks

Manufacturing is the theme of the national economy,with the introduction of Made in China 2025,the manufacturing industry is in urgent need of reform in the production process as well as detection means,the traditional detection method is mainly manual visual inspection,can not achieve 24 hours online quality inspection,which to a certain extent will lead to slow work progress.As intelligent inspection becomes the trend,it can effectively overcome the shortcomings of traditional manual work.Stamped parts are a typical component with varying process difficulties,but as the production speed increases,so does the probability of defects,for which rapid defect detection is essential.In this paper,the combination of image processing methods,image defect feature extraction algorithms,and defect area identification algorithms is used to achieve an in-depth study of the classification of defects and quantification of defect areas on stamped parts.In terms of fast recognition of surface defects on stamped parts,a method for fast detection of surface defects on stamped parts based on residual neural networks is proposed.In the methodological study of this paper,an experimental device for automated image acquisition was first designed and relevant parameters were calculated.Images of seven categories including pits,porosity,scratches,burrs,missing holes,stains and normal parts were acquired and pre-processed;a residual neural network with a depth of 50 layers was built as the backbone network to extract defect features and a neural network for region recognition Fast-RCNN was constructed as a backbone network with a depth of 50 layers to extract defect features and a Fast-RCNN for region identification,which enables the identification of defects and the labelling of defective regions on stamped parts with a small overall network width and few parameters.The obtained defect images are then binarised with global thresholding,three channels are extracted for binarisation comparison,the binarised image of the optimal channel is selected and then the binarised image under the standard image is subjected to fast convolution calculation,using different convolution kernels to determine the location of the defect area and the number of defect pixels,and the defect area is calculated based on the minimum pixel size.Finally,Labview software is used to create an operator interface for the control of the device and the real-time display of defect information,which facilitates control and monitoring and reduces learning costs.The experiments were carried out by first completing pre-processing of the obtained images to obtain a standardised image size,then making a dataset and dividing the training set and test set,and then conducting feature training and testing for a single category of defects respectively,where the recognition accuracy of burrs and scratches was low at85.35% and 87.42% respectively,and the recognition accuracy of the stain category was better at 92.33%.Finally the full dataset was then tested with a mixture of training and the average accuracy obtained was 89.55% for the seven categories of samples.