Physically synthetic data for deep learning-based visual scratch inspection of aerospace alloys with complex geometries

Aerospace alloys often operate under extreme conditions. Accurate segmenting defects in images of aerospace components is the key to quantifying the defects and evaluating their impact for part lifespan. The components are usually with complex free-form surfaces, leading to uneven light distribution in images. The variable image presentations pose a great challenge for accurate segmentation, especially with limited data. Generative adversarial networks and other training-based methods are commonly used for image generation, but they still rely on sufficient high-quality training data. In this paper, a physical-based image generation method is proposed to create any possible scratches following the physics laws to improve the scratch segmentation capability with limited data. First, an efficient scratched blade surface image generation pipeline is developed. Then, a systematic strategy to maximize the effect of physical synthetic scratch images is presented. The experiments show that the segmentation IoUs could be improved from 0.66 to 0.83 with only 20 real images for training, and reveal the influences of network structure, image and label quality, data fusion strategy on segmentation performance.