ABSTRACT
The paper aims to present an AI-based Automated Optical Inspection (AO’) software for both digital and wooden industries developed within the EdgeAT project. Current approaches rely on centralized solutions, where the computation is performed inside the inspection machine itself. Instead, we present algorithms that work at the edge to give rise to competitive solutions to existing ones. In particular, we experiment with two different tasks of defect identification: detecting the defect position within a wooden panel by using YOLO, in which a 96% accuracy is reached. Secondly, concerning the digital industry, we perform a two-step classification between defective and non-defective microchips and then between four possible defect classes in their surface exploiting a ResNet network and obtaining a 97% accuracy. We also exploit explainability tools to understand which parts of the images caused the model’s decision. After developing the AI models we port them to two less power-consuming edge devices, Nvidia Orin Nano, and Nvidia Orin AGX, observing unchanged performance.
