Research On Surface Defect Detection Technology Based On Machine Vision

With the rapid development of computer technology and automation technology,machine vision based surface defect detection technologies provide essential technical for the detection and recognition of product surface flaws.Defect detection has attracted wide attention from government and enterprises,which has been extensively employed in many civil fields,such as strip steel,rail,liquid crystal panel,textiles and so on.In this paper,we deeply study the basis of image processing and machine learning methods,and explore the recently proposed deep learning technologies.According to image processing technology,surface defect image can be roughly divided into non-pattern background,pattern texture background and pattern complex background from the background angle.Based on this,the research on surface defect segmentation mainly focus on the non-pattern background based methods and the surface defect detection of the pattern texture background based methods.For traditional machine learning technology,the research concentration is basically on defect feature extraction,feature selection and pattern classifier design,then the extracted features and classifiers are used to identify the types of surface defect images of strip steel.For deep learning,the surface defect recognition method based on convolution neural network is mainly explored.The main contributions of this paper are as follows.Firstly,a non-patterned background based surface defect segmentation method,combining Shannon entropy with improved cuckoo optimization,is proposed.Considering the simplicity of non-patterned background,the global threshold algorithm could be used to distinguish the defects.Therefore,in this paper,the simplest but very effective Shannon entropy,measuring the mutation of the defect target and the background region,is used to determine the optimal threshold,which is denoted as the optimal segmentation threshold.In order to further shorten the time of searching the optimal threshold,the cuckoo optimization algorithm proposed in recent years is used to accelerate the threshold parameter selection,which is achieved by maximizing the sum of Shannon entropy of the defective target and background area.The traditional cuckoo optimization method is improved so that the step size and the host birds'probability of finding the alien birds adaptively update which can improve the adaptability of the algorithm.Experimental results show that the proposed method perform well on image segmentation of various non-patterned background surface defects,including metal,wood,steel plate,glass and film,etc.Besides the optimal threshold can be obtained exactly,and the convergence speed is faster.Therefore,the proposed method obtain a superior performance.Secondly,a patterned fabric defect detection based on LGD and low-rank decomposition is proposed.Firstly,the LGD feature is constructed by fully combining the texture description ability of Log-Gabor filter and the accurate positioning performance of DAISY descriptor.Secondly,the LGD feature is decomposed by low-rank sparse decomposition model into sparse component and the low-rank component.The sparse part denotes the defective part,which is recovered by L₁ norm.Finally,simple threshold segmentation algorithm is applied to binarize defect regions and background regions.Compared with four recently proposed methods under four criteria,containing ROC?Receiver Operating Characteristic?curve,PR?Precision-Recall?curve,F-measure and Mean Absolute Error?MAE?.The experimental results show that the proposed method work well in surface defect detection of various patterned fabric,including star-patterned fabric?box-patterned fabric and dot-patterned fabric,achieving higher positioning accuracy and more complete description of defect shape.Thirdly,a noise robust recognition method for strip surface defects based on fusion features is proposed.Completed Local Binary Patterns?CLBP?descriptor improve the ability of local Binary Patterns?LBP?descriptor to discriminate different types of surface defects and represent fine texture details of images.Rotation Invariant Binary Gabor Pattern?BGPri?describes image shape,structure and coarse-scale texture features.Their advantages are complementary.Before fusing the two features,principal component analysis?PCA?is exploited to reduce the dimension of features,removes redundant information.Nearest Neighbor Classifier?NNC?and Random Forest?RF?classifier are used to classify defective features respectively.The proposed method is tested on NEU datasets.Comparing with four the state-of-the-art recognition methods,the experimental results show that the average recognition accuracy of each method is higher without adding gaussion noise,but the proposed method is the highest.In the case that different SNR gaussion noises are added,the average recognition accuracy of other methods decreases significantly when SNR is less than 30 dB.However,the proposed method can still maintain higher recognition accuracy,and the proposed method is time-consuming cheaply.Finally,a method of surface defect recognition based on residual network is explored.Considering that the traditional machine learning method can not achieve an end-to-end recognition,the recognition accuracy depends heavily on the extracted features and the designed classifier,resulting in poor generalization ability for new tasks.The recognition method based on convolutional neural network can perfectly solve the shortcoming of traditional method,which provides a new idea for surface defect recognition.In order to solve the problem of over-fitting caused by fewer samples,migration learning and fine-tuning techniques are used to initialize the parameters of the model through the parameters of the pre-training network.The final classification layer of the original ResNet-50 network is replaced by a layer of Softmax classification.We fix the parameters of all the previous layers,and only fine-tune the parameters of the last classification layer under the training data set.Experiments tested on the NEU dataset and the Lumber dataset show that the proposed method can efficiently combine feature extraction with pattern classification process to achieve end-to-end recognition,obtaining higher recognition accuracy on both data sets and faster convergence.In addition,the proposed method presents the superiority as compared with the LGD feature descriptor and the recognition method of the random forest classifier proposed in Chapter 3,the recognition method proposed in Chapter 4,and the recognition method based on VGGNet-16.