Defect Detection Method Based On Siamese Networks

In quality control,surface defect detection is an important technical means to ensure the quality of production.Due to the increase in production requirements and production technology,most traditional defect detection technologies have been unable to meet production requirements.In this thesis,a product surface defect detection algorithm combining saliency detection and deep learning is proposed..The main research object of this thesis is defect detection on the surface of products.The traditional method based on defect detection has a long cycle and poor versatility.With the development of deep learning,convolutional neural networks began to be applied to detection and classification.Due to the scarcity of defective samples,a large amount of data could not be provided for learning training,and the unknownity of defects caused the labels to be unknown.In order to solve these shortcomings,the article uses the significance detection,which can quickly detect the defects in the image,but the result is a collection of real defects and pseudo defects,so it is necessary to classify the authenticity defects.Due to the scarcity and unknownity of the sample of defects,this thesis studies a similarity discrimination model to replace the traditional classification model.The training of the model uses the input of the graph,which can expand the training set and solve the problem of sample scarcity.In this thesis,the twin convolutional network is selected as the similarity discriminant model,and the convolutional neural network is used as the mapping function to complete the image feature self-extraction.This method can solve the problem of small sample training well and weaken the category label without knowing the truth.The specific category of the defect.In the verification phase,the test set is divided into two parts: the painted and printed product surface image and the DAGM2007 data set.The method of this thesis is compared with the pattern-based method and the classification-based method.The test results show that the detection method of the article method is better.This thesis proposes a new method combining the saliency method and deep learning to study the coated products and printed products.The method idea can be applied not only to two-dimensional image defect detection,but also to the task of defect detection and anomaly detection of three-dimensional object images.Can be used as a new method in the field of defect detection.