Research On Surface Image Defect Detection Method For Additive Manufacturing Parts

Additive manufacturing technology is increasingly widely,which is used in high-end manufacturing fields such as aerospace manufacturing and medical equipment.Nondestructive detection technology for quality monitoring of molded parts is increasingly demanding for its accuracy,real-time performance,and ease of operation.In the paper,research is carried out on the image defect detection technology of additive manufacturing parts.Based on the actual needs of part defect detection during manufacturing,leading-edge computer vision and deep learning algorithms are introduced to solve the technical problem of real-time detection of process defects during the layer-by-layer stacking molding process.A series of defects detection methods for additive manufacturing parts applied in different scenarios are proposed,and experimental research is conducted on the proposed detection methods.The feasibility of the proposed method is quantified by using defect detection performance indicators.By analyzing the advantages of comparative research methods,it provides a research benchmark combined with image visual inspection methods for the field of defect detection of additive manufacturing parts.The main contents are summarized as follows:1)Defect detection for additive manufacturing parts using image processing method: Based on the pixel threshold difference between the image defect area and the normal area on the surface of the built part,the improved Otsu's global optimal threshold segmentation algorithm is introduced;an improved mark-controlled image watershed segmentation technology to extract defect foreground mark and background mark information to achieve image defect segmentation.The method has an accuracy rate of 95.1% for process stomatal defect detection;the wavelet transform defect detection algorithm is proposed,which use wavelet transform to extract the defect image edge feature to obtain the defect contour.The accuracy of the method for detecting process stomatal defects is 97.1%.2)Defect image classification method for additive manufacturing parts based on deep convolutional neural network: In order to be able to classify and identify defects in different processes,a lightweight defect classification model,Defect Net,was constructed using deep convolutional neural network.Multi-class and multi-label classification of part defects was realized through making additive manufacturing parts defect classification data sets.The model training accuracy rate is 99.68%,and the verification data set recognition accuracy rate is 99.05%;and two types of deep residual networks based on the Inception module are designed.Defect fracture image was obtained through the tensile test of the fracture of the additive manufacturing part,and the classification performance of the defect image was studied based on the designed network model.The classification accuracy rate reached 98.2%,the recall rate was 96.6%,and the accuracy rate was 97%,which meets the performance requirements of industrial-grade defect identification solutions.3)Semantic segmentation method for defect images of additive manufacturing parts based on full convolutional neural network: Aiming at the problem that image processing algorithms cannot implement part defect classification and complete pixel-level segmentation,a residual attention full convolutional network model is proposed.The multimodule cascade coupling method realizes pixel-level classification and segmentation of image defects.At the same time,the semantic defect segmentation data set was collected and annotated,the data set was used to study the defect detection performance.The model training accuracy rate reaches 99.89%,and the verification accuracy rate reaches 97.30%.The average DSC coefficient of various defects is 90.82%,and the average AS similarity of all defects is 94.41%.4)Defect image instance-level segmentation method for additive manufacturing parts: Aiming at the problem that the defect image semantic segmentation method cannot complete the difference recognition of the same type of defects,further research proposed a defect image instance segmentation method using three deep neural networks modules.The network first extracts the image features,and then identifies the defect features in the image to obtain the target defect area.Finally,the extracted defect areas are classified and the area mask is segmented to implement the defect instance segmentation of additive manufacturing parts.Experiments show that the accuracy rate of predicting 6 types of defects is 95.90%,and the precision rate is 91.50%.The instance segmentation performance index m AP(mean defect recognition average AP value)is 89.40 in the defect instance segmentation task of this research.It is verified by experiments that the performance of defect detection is higher than the three types of mainstream object detection algorithms.At the same time,the segmentation speed of the defect instance reaches 12 frame images per second,which can realize real-time defect detection during the manufacturing process.