Research And Implementation Of Appearance Defects Small Sample Vision Detection Method For Injection Molding Parts

At present,the detection situation of the appearance defects of injection-molded parts in domestic enterprises is mainly by manual inspection,some adopt machine vision inspection based on feature engineering and little deep learning detection.However,the manual detection method is inefficient and the quality control is unstable;the feature engineering method relies too much on the artificial design features and has poor flexibility.The deep learning method has the ability to automatically learn and extract features,but this detection method has a long way to go from being widely applied for actual production.The main reason are that the method is driven by data,the demand for label data samples is extremely high,which is contradicting with the situation of lacking high quality samples data.The small sample convolutional neural network small sample detection method is the future of the appearance inspection.Therefore,this paper has conducted in-depth research on the injection parts appearance detection system and the small sample convolutional neural network detection method.The main work are as follows:(1)Combined with actual production designed an inspection process,and built an automatic defect detection system for injection molded parts.The hardware system of the inspection platform was designed,and the industrial camera,lens and light source were well picked up,and reasonable lighting and image acquisition schemes were designed.Based on the modular design idea,the module structure of the software system was designed and various functional modules has developed based on the.Net platform,which achieved automatic detection of defects in the appearance of injection-molded parts.(2)In order to solve the problem that the feature engineering method relies too much on the artificial design features and performs poor flexibility,the convolutional neural network method was applied to perform automatic feature extraction and defect detection on the injection-molded parts.The specific structure of the convolutional network was designed,the training parameters and training set size of the network are optimized experimentally,and the uniform detection method of the appearance defects of the injection-molded parts is realized.(3)In order to overcome the dependence of the convolutional neural network detection method on a large scale of data samples,migration-learning method was adopted.The network training scheme for model migration was designed,and the key parameters such as the number of layers to be migrated and the number of layers to be fixed was determined by experiments.The model of defect detection of injection-molded parts based on model parameters migration was constructed.(4)The performance of the defect detection model was verified by experiments.Firstly,multi-classification experiments proved that both of the convolutional neural network model and the migration model has a good correctness and flexibility in detection.Then,the important role of the migration model in reducing the need for data samples was proved by experiments.In the end,experiments show that both the convolutional network model and the migration-learning model performed well in various defects detection of the earphone shell products,and achieved above 97% of detection accuracy.At the same time,the migration model reduced the demand for data samples by 68%,which has a high application prospect.