Image Synthesis of Metallic Surface for Training Dataset in Visual Inspection

In this article, the authors observe the issue of the lack of quality data for training of the automated visual inspection systems. Given the narrow specialization of such systems and the absence of comprehensive open-access datasets containing images of metallic surfaces with various defects, solving this problem becomes a highly relevant task. The authors propose three methods for synthesizing images in the absence of real metallic samples: manual creation using the Blender 3D modeling environment, procedural generation using the Unity game engine, and training a custom core model for Stable Diffusion to subsequently generate images using text prompts. The authors outline criteria for comparing the implemented image synthesis methods and provide recommendations for their further use.