Contour Segmentation of Image Damage Detection Based on Fully Convolutional Neural Network

Abstract

Damage detection is a critical task in monitoring and inspection for aircraft internal structures. In the actual situation, most were nondestructive evaluation, such as ultrasonic inspection, which scan the internal structure of the aircraft, to obtain the damage inside for the testing parts. However, there is still no accurate standard for damage assessment and quantification on the scanned images by ultrasonic, due to the low image resolution, or the complicated scan result. The traditional contour detection algorithms, such as Canny Edge Detection (CED), color threshold, are difficult to apply on the damage contour segmentation for such images. In view of the progress of deep learning methods, the current study proposes a damage detection method based on Fully Convolutional Network (FCN), for the contour segmentation on ultrasonic detection of damage images. The whole FCN network for contour segmentation with the Visual Geometry Group (VGG) based is trained end-to-end on a set of 2000 256 × 256 pixels damage-labeled scanned images of a certain alloy which can be made for fan blade, another 400 images are used to test the FCN method. The contour extracted by FCN are qualitatively similar to the ground truth, achieve over 92% average precision. The FCN performance is better than the traditional algorithm, and the training model can be used for transfer learning to adapt to the extraction of different damage types. The results of segmentation can be further used for quantitative analysis of damage area.