The Research Of Coating Defeet Detection Technology Based On Deep Learning

With the rapid development of image sensing techn ology,visual inspection is widely used in various fields as a non-contact sensing technology,which can replace manual labor to do automatic product defect detection and quality verification,and even meets the requirements of automatic control of the entire production process.At present,most of the online detection processes for defects of non-woven fabrics rely on manual 'visual inspection.The detection speed and efficiency are slow and human factors have great influence on the detecting results,so real-time rapid detection cannot be realized.The use of machine vision methods for surface defect detection,most of which have complex artificial feature design,and the accuracy is not high.In this paper,a deep defect-based coating defect detection system is designed.The high-resolution industrial CMOS camera is used to obtain the coating defect image on the coating line,which is divided and normalized into 256*256 size.It preprocesses the image through image enhancement and noise preprocessing and inputs the preprocessed sample image into the convolutional neural network,and automatically extracts the feature vector through the iteration of big data.For the parameters in the training network,the analysis method is used one by one to determine the reasonable network parameters and hierarchical structure,and multiple experiments are carried out to determine the validity of the parameters.Compared with the traditional machine vision SVM detection method,the superiority of the defect detection method using convolutional neural network is shown.The paper uses th,e Linux system to build a deep learning framework Caffe for model training,and the experimental verification on the HiSilicon Hikey 970 hardware platform,effectively achieving the defect detection of the coating production line.Through training 5,352 sample data of wrinkles,vertical lines,bright spots and normal images were trained and 1200 samples are tested.And the defect image is rotated,mirrored and stretched during the training process to increase the number of sample libraries to improve the overall network generalization.The experimental results show that the defect detection system can operate stably with accurate recognition rate of 91%and recognition rate of 0.91s.It has achieved certain research results and can be applied to the same coating manufacturing field,which has certain reference significance for subsequent research.