Research On The Detection Of Defects In Metal Lattices Structure Based On YOLOV3 Algorithm

The metal lattice structure manufactured by selective laser melting technology has a great advantage in light weight,strength and functionality compared with traditional metal materials due to its "hollow" characteristics.It has attracted the attention of researchers as a potential lightweight material.Lattice structure is often used in aerospace,equipment manufacturing and other military fields.In the manufacturing process,defects will inevitably occur,which will seriously reduce the performance and reliability of the structure and unable to meet the high-performance requirements of the structure system.At present,for the defect detection of the lattice structure,the defect detection is mainly performed by manually reviewing the CT images.However,there will be problems of inefficiency and missed detection due to personal emotions and fatigue.There are few reports on the automatic detection of defects and the mechanism of intelligent review,and efficient and standardized detection methods have not yet been formed.This paper proposes a method for detecting lattice structure defects based on deep learning algorithms,which realizes the classification and localization of defects,and provides a new method for the field of metal lattice detection for lattice structure.Aiming at the problem of defect detection for lattice structure,firstly combine industrial CT equipment and 3D reconstruction software to obtain slice images of lattice structure defects,and use data augmentation methods to increase the number of samples,and then label the slice images to establish defect data sets.To solve the problem that the defect in lattice structure is difficult to detect due to its small size,weak defect features and high similarity with the background,a detection method based on YOLOV3 algorithm is proposed.This method takes advantage of deep learning network model in feature extraction,uses multi-scale prediction network for feature fusion,and treats the classification and location of defects as regression problems.In order to improve the detection speed,the feature extraction network is compressed.Through sparse training,the BN layer weights are induced by regularization,and the BN layer weight gradient is changed,making the model more sparse.Under the premise of notexceeding the threshold,the network model is pruned at different proportions.Finally,the new model is analyzed based on indicators such as map and network forward inference speed,and the effects of different pruning ratios on the model are compared.