Page 159 - Read Online

Wang et al. J. Mater. Inf. 2026, 6, 13 Page 3 of 19

Figure 1. Various aero-engine blades with free-form surfaces. Photograph taken by the authors.

Figure 2. Variety of scratches presentations in images. They may look bright, or dark, and often continuously change at different areas on
complex surfaces.

Physical-based image generation relies solely on simulating the lighting and imaging processes to generate
images using computer graphics. Because it can simulate all possible imaging conditions and complex
geometries without being limited by existing datasets, it has recently emerged as a promising solution for
defect image generation. The procedures for defect modeling in AVI have been developed [26-30] , and the
performance of physical synthetic data for AVI across various objects has also been evaluated [31,32] . These
studies demonstrate that the approach works well in certain defect detection scenarios. However, most
studies have focused on image classification or object detection. How to define scratches, generate, and
utilize physical synthetic data for high-accuracy scratch segmentation on aerospace alloys with complex
geometries still requires further investigation.

In this paper, we introduce the physical synthetic data into the surface inspection of aerospace alloy scenarios
and investigate its influence on the performance of defect segmentation. Aero-engine blades, among the
most complex parts made from aerospace alloys, are selected as the verification object, and scratches, one of
the most common defect types on aerospace components , are chosen as the defect type. First, an efficient
[1,5]
pipeline for generating scratch images on free-form surfaces is developed to quickly produce synthetic defect
images. Then, an application strategy for using physical synthetic data in scratch segmentation model
construction is proposed, based on trends revealed by a series of experiments. The main contributions of this
paper are:
• Introducing physical synthetic data for scratch segmentation of complex aerospace alloy parts.
• Developing an efficient image generation pipeline, with results showing that synthetic data can evidently
improve segmentation performance with limited real data.