CenterNet-based defect detection for additive manufacturing

Abstract

Additive manufacturing (AM) has been widely used in the fabrication of optical components. However, surface defects generated during the AM process have an adverse effect on surface quality. Although some studies have explored the defect features based on the processing of information including images, acoustic signals, thermal history, etc., they focus mainly on defect classification or one type of defect detection. Over recent years, convolution neural networks have displayed promising performance in object detection in images in various fields. Therefore, in this paper, to detect and characterize surface defects more comprehensively and accurately, a novel defect detection model based on CenterNet is presented to extract the defect features, including type, location and count simultaneously, in which there are four output heads to predict heatmaps, object size, local offset, and density map, respectively. Moreover, count loss is added in the original objective function to boost the detection performance. To perform the model validation, surface defect dataset is captured through scanning electron microscope on the surfaces of the workpiece made of 316L fabricated by AM. A series of experiments was conducted and the proposed model achieved better detection accuracy on defect dataset compared with other state-of-the-art models.