Research On Key Algorithm Of Product Surface Defect Detection Based On Machine Vision

In recent years,developed countries have started to upgrade their manufacturing industry,mainly represented by German industry 4.0.Manufacturing industry is not only the pillar industry in the national economy,but also the main battlefield to promote innovation and industrial upgrading.In the strategy of "Made in China 2025",it is pointed out that the quality of products should be improved faster and the implementation plan for improving product quality should be formulated.For an enterprise,product quality is the cornerstone of enterprise development and competition.Improving the quality of products is very important for enterprises.Therefore,in many industrial fields,product surface defect detection has attracted a large number of researchers to conduct in-depth research on it.To further improve the automation of product defect detection,using machine learning or deep learning methods has become the development direction.In the industrial production process,manual measurement and judgment will produce errors and errors due to fatigue,individual differences of operators.Visual saliency method to simulate people for defect detection of products can increase efficiency and accuracy,and can easily distinguish defective areas from non-defective areas.Therefore,the research of defect detection algorithm based on machine vision plays an important role in the development of intelligent production lines,and is also a key step in the intelligent manufacturing of industrial production in China.Thus,it is of great practical significance to study the application of machine vision in product defect detection.The research of this topic has important theoretical significance and practical value for the realization of intelligent manufacturing.The specific research contents are as follows:1.An image segmentation surface defect algorithm based on saliency detection of regional features clustering is presented.The traditional machine learning based surface defect detection algorithm has low accuracy and weak robustness.This paper presents an image segmentation surface defect algorithm based on saliency detection of regional features clustering.First,the image is segmented using a super-pixel algorithm,and then the eigenvectors of the segmented result are computed.Then the feature vectors are clustered using a multibandwidth parameterless clustering algorithm,and the saliency maps calculated under ten bandwidths is obtained.Saliency maps are merged using a neural network.Add the resulting saliency map as a new feature to the eigenvector and repeat the process again.The final defect image segmentation map was obtained after ten repetitions.Experiments show that this algorithm improves the performance,accuracy,recall,F score and MAE compared with other algorithms.2.A learning algorithm based on U-net and transfer learning is presented.For the current deep learning image segmentation algorithm,there is still a lot of space to improve the accuracy and reduce the amount of calculation.Using transfer learning or improved algorithm can improve the learning accuracy and achieve good results.Combining batch normalization(BN)layers into convolution layers can reduce the amount of calculation.First,extend the dataset using data augmentation.Efficient Net is then used as the backbone network to extract features.Use U-net to fuse features at multiple scales.Finally,four types of defects,size 512x512x4,are obtained by a layer of convolution.Experiments show that the algorithm has higher accuracy and robustness than other algorithms.The average Dice coefficient is 0.905.3.An algorithm based on transfer learning and image segmentation is presented.First,the data set is expanded using data augmentation technology.Next,we use Efficient Net as the backbone to get the feature map.Five different scales of feature map are obtained from the repeated Bi FPN.Then the five feature maps are fused together by deconvolution,and a layer of convolution is used to get the masks of various defects.After the final training,we use batch normalization and convolution layer fusion to reduce the running time of the algorithm.Because the mish activation function can get better gradient without saturation,the activation function in this paper uses mish,which makes the model more accurate and has better generalization.The experimental results show that the average Dice coefficient of the algorithm is 0.912.Defect detection of steel surface is better than other deep learning algorithms.4.A surface defect detection algorithm combining classification with object detection model is proposed.The classification method based on deep learning can only classify the image,but cannot determine the location and size of the defect.This has a great impact on the future data analysis.Based on GAN and reinforcement learning,it is difficult to train a stable and accurate model.This chapter realizes the automatic detection and location of steel plate surface defects,further improves the accuracy and stability of the algorithm,and reduces the average running time of the algorithm.Through the analysis,it is found that the surface defects of pits are usually small and not obvious,so the accuracy of such defects is not high.In the improved Faster R-CNN model,FPN and SPP are added to the feature extraction part of Faster R-CNN to improve the recognition accuracy of this kind of defects.In addition,the accuracy of crazing defect detection is not high,because it is narrow and long.By changing the size of the default anchor of Faster R-CNN,the positioning accuracy is improved.In the improved Resnet50-vd model,which is the backbone of the classification model and the target recognition model,DCN and improved Cutout data augmentation are added to better detect various shapes of defects,with higher accuracy and robustness.Finally,the accuracy of classification model and object detection model is analyzed,and the accuracy of the whole model is 0.982.