Research On Surface Defect Recognition Based On Feature-cluster Based Convolutional Neural Network

Surface defect is a common quality problem,and it widely exists in metal,ceramic,wood,textile,and other products.Those defects influence the product appearances and performances,which might cause unnecessary losses and safety loopholes.Thus,how to recognize defects accurately and effectively is an urgently needed problem.Recently,deep learning(DL)has developed rapidly,and using convolutional neural network(CNN),one of DL models,in surface defect recognition has driven more and more attention.However,the previous work still has several limitations.From the perspectives of CNN models,the existing methods lack analysis of the training process of CNN,so it is difficult to guide the training process.From the perspectives of defect recognition problems,several tasks,such as small-scale-sample and low-quality defect recognition tasks,have not been addressed.To overcome these limitations,this paper proposes a feature-cluster based CNN(FCCNN).In FCCNN,a feature cluster mechanism is introduced into CNN to guide the training and improve the recognition results.Based on these improvements,FCCNN has been developed into surface defect recognition tasks,and improved for small-scale-sample and low-quality defect recognition problems.Finally,the proposed method is developed into a real-world defect recognition case.The main contributions of this paper are as follows:For CNN models,by analyzing the feature space,a feature-cluster based CNN(FCCNN)is proposed,a network structure with dual outputs of feature and recognition label is designed,and a feature cluster mechanism is introduced to guide the model training.The experimental results based on public benchmarks,such as MNIST and CIFAR10,indicate that FCCNN has outperformed the other methods.These results suggest that the improvements in FCCNN are effective.For conventional surface defect recognition problems,a surface defect recognition method based on FCCNN is proposed.In this method,a VGG16-based network is designed,and a pretrained weight is used to replace the pretraining step and save the computation cost.Furthermore,this method is also improved by dropout and global average pooling for better recognition results.The experimental results based on steel surface defects and fabric defects indicate that the proposed method has better recognition results.Comparing with the other defect methods,the recognition results are improved by the proposed method.For the small-scale-sample defect recognition problems,this paper proposed a multiscale feature fusion-based FCCNN.The proposed method uses a Gaussian pyramid to extract multi-scale defect feature.After that,the proposed method builds several FCCNNs for each scale individually,and fuse these models to improve the recognition results of small-scale-sample defects.The experimental results based on steel surface defects and wood surface defects indicate that the proposed method has an outstanding performance.Comparing with the other defect recognition methods,the recognition results are improved.These results suggest the improvements of the proposed method is effective.For the problem of low-quality defect recognition,a low-quality defect recognition method based on FCCNN and image reconstruction is proposed.In the proposed method,a generative adversarial network(GAN)is improved to learn the potential feature information of the defect images,and to reconstruct high-quality defect images from the low-quality ones.After that,the proposed GAN can reconstruct high-quality images and fill up the lost information.Based on the reconstructed images,an FCCNN is used for defect recognition.With these improvements,the proposed method can reduce the influence of low-quality defects for defect recognition.The experimental results based on low-quality defect images,involving noisy images and masked images,indicate that the proposed method can reconstruct high-quality defect images,recognize them accurately.Comparing with the other defect recognition methods,the recognition results are improved greatly.Based on the research above,the paper develops the proposed FCCNN into a realworld defect recognition case.This case is from a cold-strip steel workshop in China.With the analysis of the background and characteristics,several adjustments are implemented into the proposed FCCNN.The experimental results indicate that the proposed method outperforms the current recognition method in this workshop.This result suggests that the proposed methods in this paper have good potential for application.Finally,this paper summarizes the contributions and innovations,and discusses some research direction in future work.