Research On Visual Inspection Method Of Appearance Defects In Injection Parts Sorting System

With the continuous progress of modern industrial manufacturing capacity,Plastic molded parts is the main production method of plastic injection parts are widely used.Injection molded parts may produce some appearance defects during the injection molding process,and these defective injection molded parts cannot be used for subsequent product assembly.Therefore,after injection molding,the appearance defect detection must be carried out.At present,manual visual detection is still the main detection method of injection molding parts production enterprises.This traditional manual detection has many shortcomings.The missed detection caused by human subjective negligence is unavoidable,and the stability of detection is difficult to guarantee.Traditional machine vision detection methods based on machine learning mainly select specific features for specific detection targets as classification reference.This detection method usually has a strong correlation with the characteristics of the inspected product and the inspected defect,and requires high engineering experience of the designer.The detection accuracy is highly dependent on manual experience.Based on these problems,the application of convolutional neural network algorithm to the defect detection of injection molding parts was studied in this paper.In view of the traditional image recognition technology of machine vision technology existing engineering experience,high detection accuracy is difficult to guarantee some shortages and poor product suitability,injection molded parts to make use of convolution neural network recognition method on the one hand can realize the plastic automatic defect inspection,on the other hand can also overcome the problem such as the limitation of the traditional characteristics of the engineering algorithm,has important research significance.Therefore,first of all,according to the appearance defect characteristics of injection molded parts,an appropriate image acquisition device was selected and an image acquisition device was built.The image data set of the training set was obtained by the image acquisition device built,and the data set was processed.Then,according to the characteristics of the small plastic appearance defects,based on the analysis of the characteristics of the three classical convolution neural network,a combination of three network in convolution kernel size and network layer in terms of their respective advantages,determined the suitable for injection molding parts surface defect detection based on convolution neural network structure,has obtained the certain effect.Convolution based on neural network in the whole connecting layer is the increase of network parameters of the problems of large amount of data,and further puts forward using global average pooling instead of full connection layer,reduce the network parameters,the experiments show that the improved convolution neural network on the identification accuracy is still maintained a 97% effect,recognition time increased nearly 50%,simplifies the network computation...