Traceability and uncertainty of defects automated measurements by CNN-powered Machine Vision Systems

Surface geometric imperfections can be automatically inspected by machine vision systems. State-of-the-art applications prefer resorting to image analysis by Convolutional Neural Networks (CNNs), rather than traditional traceable inspection methods. CNNs have the advantage of greater speed, flexibility and automation but lack traceability, thus hindering quantitative quality controls and tolerances verification. This work proposes a methodology to estimate the uncertainty of automated measurements of surface geometrical imperfections based on CNNs while establishing traceability by leveraging on a photogrammetric system. The methodology is demonstrated on a gas metal arc welding of aluminium alloys for inspecting and measuring the quality of surface pores.