Research On Unsupervised Defect Detection Algorithm Of Product Surface

In industrial manufacturing and production,due to some factors,some defects may appear on the product surface,and the sensory level of the product surface also affects the development of the whole industry.In the traditional defect detection of product surface,the template of product sample should be established first,and then the image of the sample to be tested should be compared with the standard template.In the acquisition process,due to the influence of light intensity,acquisition location and other factors,it is necessary to carry out pre-processing,template matching and other operations before comparison,which makes the whole detection time-consuming and labor-consuming.In recent years,the development of deep learning has made a great breakthrough in supervised learning tasks such as target detection,semantic segmentation,etc.a prominent feature of deep learning is that the network can automatically learn image features.Compared with the traditional methods,most of the defect detection based on deep learning is supervised learning,which can improve the detection accuracy to a certain extent,but it needs a large number of defect sample data for training.In the production process,affected by various random factors,it is impossible to fully predict the type of defects.Therefore,using deep learning for supervised network training learning is not suitable for defect detection task.In view of the problems of the above methods and the needs of practical application,this dissertation proposes a product surface defect detection method based on unsupervised learning.The main contributions are as follows:A residual self coding image reconstruction model based on depth convolution network is proposed.In this dissertation,for the data set with non repetitive periodic background,the image reconstruction is based on self coding and residual coding decoding network.The structure of coding and decoding is symmetrical,and the structure of overlapping feature map is designed in the coding network to retain more image feature information.Using the same type of data set,the network actively learns the feature vector representation of the data in the hidden space to get the corresponding reconstructed image.For the sample image to be tested,if it conforms to the image distribution of training set type,the network can reconstruct it;if it is other kinds of data,the feature vector obtained by coding is quite different from the feature vector distribution learned in the training stage,so it cannot be reconstructed well.The reconstruction experiment on the cigarette box printing data set shows that the residual self coding network model based on the deep convolution neural network has a certain improvement in the accuracy of image reconstruction.Based on the reconstruction model,an unsupervised learning algorithm for product surface defect detection is proposed.Aiming at the tedious work of traditional detection methods and the requirement of supervised detection algorithm in deep learning for the number of bad product data sets and defect types,an unsupervised defect detection algorithm based on image reconstruction is proposed.Only using the good product data,and without knowing the type of defect and the number of defect data,add coding settings after the self coding network of image reconstruction,and then encode the reconstructed good product image for the second time.In the whole model,the image is encoded once and twice.If it is good product data,the difference between the feature vectors obtained after the two times of encoding will not be too large.If it is defect data,because the network only learns the good product data,the difference between the feature vectors obtained after the two times of encoding is large.By setting the corresponding thresholds at the feature vector level and image level,complete the judgment of defect data in the detection stage.