Visual Quality Control via eXplainable AI and the Case of Human in the AI Loop

Vision-based quality control is commonly employed in produced assets and consumer goods. Typically performed in the past through human manual inspection, it is steadily being replaced by automated inspection systems employing some form of data-driven artificial intelligence to assign products to passing or not the quality control. Success has been reported in the literature on employing deep learning approaches for imagebased classification. The motivation for such choices stems mostly from the ability of deep learning to develop layered feature representations of input stimuli, a process loosely aligned with human capabilities to develop hierarchical perception abstractions.
However, several challenges remain, including the lack of generalisation capabilities of such models and that outcomes are rarely understood and trusted by humans. This paper explores the second challenge on an industrial case of image-based quality inspection. It reports results from applying both a posteriori and a priori explainability as a step towards involving the human in the AI Loop.