Surface Defect Detection Algorithm Towards Unbalance Sample

China,the world first largest manufacturing power,produces lots of products every day.For some of these products,such as glass,fabric,ceramic tile,etc.,it is not only required to have usability,but also no defect on the surface of the product.With the increasing market demand,how to detect the surface of a large number of products in time and with high quality has become a research hotspot in the industrial field.Because of the limitations of traditional image processing technology in the face of complex tasks and the advantages of deep learning network in image data processing,the method based on deep learning,especially convolution neural network,has become the mainstream solution for product surface defect detection.At present,the research on defect detection includes defect location and defect multi classification.This thesis focuses on two issues.One is how to use unsupervised learning to make accurate defect region proposal,and the other is how to combine unsupervised learning with supervised learning for effective multi classification.The defect detection algorithms currently have the following challenges:1.The imbalance problem of samples leads to high false rejection rate and false detection rate of the model based on supervised learning.Obtaining a large number of annotation defect data required for training model is very difficult and expensive.2.Although the method based on unsupervised learning is not affected by the imbalance problem,its proposal for defective regions is not complete and it cannot achieve multi-classification of defects.To solve these problems and challenges,this thesis proposes a surface defect detection algorithm based on unsupervised and supervised learning for unbalanced samples.At present,the research on defect detection includes defect location and defect classification.For these two series connected sub tasks,the main method is to propose the defect region by the improved multi-layer feature convolution auto-encoder(MLF-CAE)based on unsupervised learning,and then to classify and fine tune the defect by the multi-scale detection network based on supervised learning.The main work of this thesis is as follows:Firstly,a multi-layer feature convolution auto-encoder is proposed to improve the Recall rate of defective regions,and combined with an improved two-scan connected region algorithm to produce defect region proposals.Secondly,the total variation loss is introduced into MLF-CAE,which can keep the generated image segment smooth and effectively suppress the noise of segmented image,and improve the precision of defect region proposals.Lastly,fusion the unsupervised defect region proposal network with the multi-scale detection network to form a new defect detection model,which can take into account the detection effects of defects of different scale,and use non-maximum suppression strategies to screen the final results?Finally,the thesis validates the MLF-CAE model and the complete defect detection model on the Fabrics and the aluminum surface dataset.The experimental results indicate that the performance of the proposed two models has obvious improvement compared with the previous methods and the current mainstream methods.What' s more,by changing the proportion of annotation samples in the training set,it is verified that the proposed model has better robustness in the face of unbalanced samples.