Research For Surface Defect Detection Based On Small Sample Machine Learning

In modern industrial production,appearance inspection of products is crucial in quality inspection.Therefore,it is necessary to develop powerful intelligent appearance defect detection technology.The traditional defect detection method is extracting the features of the object and designing the corresponding detection algorithm.With the development of machine learning,many powerful machine learning algorithms are gradually applied in surface defect detection and achieved good results.However,defect samples are difficult to obtain in industrial production.At the same time,when the number of training samples is extremely limited,the effect of defect detection based on machine learning algorithm is unsatisfactory.In view of this situation,the machine learning method in surface defect detection is researched in this paper.Firstly,one model containing digital image processing and Naive Bayes analysis is proposed in the situation which few kinds of defects and samples.In the learning process,the method first needs to classify different kinds of samples.Then corresponding enhancement process for different types of images is carried out.After that,the enhanced image data are extracted and quantified.After completing learning process,image enhancement and feature extraction are also carried out.After completing the process,the Bayes model is established based on the type and distribution of samples in two parts.Image information is transformed into probability information which could be accepted by the model.Characteristic information in learning process is transformed into prior probability through Bayes model.In the detection process,the posterior probability of the object to be tested is calculated by prior probability to complete the detection.Compared with machine vision methods,this model can achieve higher detection accuracy with few learning samples.The detection efficiency is higher and need less training samples.Secondly,the deep neural networks which are available can reach very high accuracy in image detection and classification.However,deep neural network needs so many training samples that can achieve high accuracy.Thus,in the case of insufficient samples,using Generative Adversarial Networks(GAN)can enlarge data sets.It can effectively overcome over-fitting and low detection accuracy caused by insufficient training samples.But conventional GAN model is hard to generate high quality training samples.Therefore,this paper proposes a Defect Enhancement GAN(DEGAN)model,which can generate high-definition and high-diversity defect data.It obtain the incentive for the follow-up training of the network by calculating the distance between real sample and generated sample.Then the output of network need to be graded.In this way,the quality of pictures generated from the network is obviously improved.Then,a defect detection model based on DEGAN and deep neural network is proposed.Experiments show that the model has higher detection accuracy and better adaptability than the deep neural network in the case of small samples.Furthermore,Combining the idea of antagonistic learning in GAN with transfer learning can also achieve outstanding performance in the field of surface defect detection.This paper proposed a migration learning defect detection network based on domain antagonism.In the case of small sample of fabric data,a transfer learning model is established to detect fabric defects with the help of a large number of magnetic ring data.The model of transfer learning network designed in this paper includes feature extraction module,defect detection module and domain classification module.In the training process,firstly,the features of target data set and auxiliary data set are extracted.Then the common features of the two types of data samples are learned by domain classifier.Finally,the defect detection module and domain classification module are used to learn the common features of the two types of data sets.At this time,the training and optimization of detection network can be carried out through common features.Experiments show that the model can carry on transfer learning by similar auxiliary data.It effectively reducing the waste of data and computing resources.Compared with deep neural network,this detection network can achieve higher detection accuracy with few sample data.Finally,the work done in this paper is summarized,and the future research directions of surface defect detection methods are prospected.