Lovro Bosnar

• Product manufacturing is performed in a massively automated and increasingly customized manner. However, overall production speed is limited by automation of inspection since each product has to ensure the required quality. A widespread and often-used quality assurance method is visual surface inspection. Automated surface inspection relies on an inspection plan and defect recognition algorithms. Both inspection planning and defect recognition algorithms development heavily rely on the availability of representative image data containing various product surface textures and imperfections showing a wide variety of possible surface responses to different viewing and lighting conditions. Due to the advancements in manufacturing, defects in products occur rarely, with different frequencies of appearance, followed by a subjective and laborious annotation process. Further, since the surface texture is often not relevant to product performance and thus not controlled, products with different surface textures are not treated as different product samples and thus not provided. Motivated by aforementioned problems, this work introduces the following contributions: (1) image synthesis requirements for industrial quality inspection and a novel realistic image synthesis pipeline satisfying those requirements (Chapter 4), (2) texture synthesis requirements for industrial quality inspection and a procedural approach to parameterized surface texture modeling incorporating domain knowledge (Chapter 5) and (3) defect synthesis requirements for industrial quality inspection as well as a procedural approach to parameterized defect modeling (Chapter 6). The contributions presented in this thesis, make it possible to obtain, in a controllable and automated manner, the required amount of image data, containing realistic and varying surface textures resembling machining surfaces as well as diversified geometrical defects with automated, pixel-precise annotations (Chapters 7,8). The presented contributions enable the inspection planning and development of machine vision algorithms for defect recognition to be performed completely virtually, by inspection planning experts, without computer graphics knowledge.