A Novel Depth-Connected Region-Based Convolutional Neural Network for Small Defect Detection in Additive Manufacturing

Abstract

Defect detection on the computed tomography (CT) images plays an important role in the development of metallic additive manufacturing (AM). Although some deep learning techniques have been adopted in the CT image-based defect detection problem, it is still a challenging task to accurately detect small-size defects in the presence of undesirable noises. In this paper, a novel defect detection method, namely, the depth-connected region-based convolutional neural network (DC-RCNN), is proposed to detect small defects and reduce the influence of noises. In particular, a saliency-guided region proposal method is first developed to generate small-size region proposals with the aim to accommodate the small defects. Then, the main architecture of DC-RCNN is proposed to extract and connect the consistent features across multiple frames, thereby reducing the influence of randomly distributed noises. Moreover, the transfer learning technique is utilized to improve the generalization ability of the proposed DC-RCNN. In order to verify the effectiveness and superiority, the proposed method is applied to the real-world AM data for defect detection. The experimental validations show that the proposed DC-RCNN is able to detect the small-size defects under noises and outperforms the original RCNN method in terms of detection accuracy and running time.