Adaptive Defect Detection for 3-D Printed Lattice Structures Based on Improved Faster R-CNN
Yuyan Zhang, Zhiwei Zhang, Kai Fu, Xiaoyuan Luo
Abstract
To detect the internal small defects of three-dimensional (3D) printed lattice structural samples, this study proposes an adaptive defect detection method based on a faster region-based convolutional neural networks (Faster R-CNN) structure. A K-medoids clustering algorithm, which adopts the Manhattan distance to calculate clustering centers, is used to adaptively select the optimal preset anchors from 3D printed lattice structures obtained via a computed tomography slices dataset. Based on these preset anchors, an improved defect detection model for over-melt based on Faster R-CNN is constructed. To improve the generalization ability, data augmentation methods are used for the computed tomography slices and a fine-tuning strategy is used for Faster R-CNN. The experimental results show that the defects of the lattice structure can be effectively detected. The adaptive defect detection model achieves the expected average precision of 93.4%. The feasibility of the adaptive defect detection method is thereby verified.