Cross-modal knowledge distillation for detecting color-related defects from monochrome images in photovoltaic cell manufacturing

This paper proposes and validates a cross-modal knowledge distillation scheme for training a detector of color anomalies in photovoltaic cells using monochrome images. A reference final-inspection model trained on three-channel RGB images serves as the knowledge source; its predictions are used to generate pseudo-labels and to automatically assemble a paired “RGB-monochrome” training set. A target detector is then trained to operate on monochrome data acquired at intermediate inspection nodes of the production line. Experiments on manufacturing line data show that, for a subset of color anomalies, diagnostic cues learned from RGB imaging remain informative in the monochrome modality, enabling reliable detection when color imaging is unavailable at the target inspection node. The practical contribution is the ability to leverage already deployed monochrome cameras for earlier identification of process deviations.