Surface Defect Detection Based On Convolutional Neural Network

Product surface inspection is an important part in industrial production.With the development of computer hardware and detection algorithms,artificial detection is gradually replaced by machine vision method.Product surface often exists stains,scratches,cracks,deformation,etc.So there are many kinds of image to be detected.Industrial detection algorithm lacks extensive transfer learning ability.A general algorithm is pursued in industrial detection.For industrial products surface quality detection,we propose a convolution neural network algorithm(CNN)on image patches with local information.Our algorithm builds a convolutional neural network to classify image blocks,and uses voting mechanism to realize localization.By our algorithm,we successfully distinguish target region and background in the image.And the location of the defect area is achieved by votes.On the basis of detection algorithm,we speed up training speed by adding a branch to network.At the same time,to solve the problem of slow detection speed,image cutting is replaced by sliding window on feature image to reduce the repeated computation in the classification process.After detection speed acceleration,we propose to use parallel network structure.Detection results of two networks use weighted method to reduce detection error rate and get finer positioning boundary.In this paper,three different texture image sets provided by the German pattern recognition association(GAPR)and two image sets with metal gaskets image and screws image are detected.In the larger three groups of texture images,our algorithm all reach detection accuracy 95%.The highest one reaches detection accuracy 98.6%.In the structural image sets of metal gasket and screws,our algorithm also reaches detection accuracy above 89%.At the same time,training speed is about 8 times faster,and detection speed is about 17 times faster with a detection time about 0.5 s.Experiment results show that our algorithm has a wide range of applicability and high detection accuracy in these kinds of image set.