Defects Detection In Rotor Composite Parts Using Instance Segmentation

Abstract

During the manufacturing process of rotor composite parts, each part is systematically controlled using a Radiographic Testing (RT) approach. This is a non-destructive testing (NDT) method, which uses x-rays to examine the internal structure of manufactured components identifying any flaws or defects within the material. In this paper, an instance segmentation approach is used to detect a specific defect in radiographic images of composites parts, based on a Mask R-CNN model. Instance segmentation represents a significant advancement in computer vision compared to more conventional approaches (such as classification or object detection). Application to an industrial case is presented here with a precision to predict "defects" of 85%, a recall of 96% and a F1 score of 90%. We also introduce an original labeling technique well suited for industrial purpose.