LEARNING SHADOW REMOVAL FROM UNPAIRED SAMPLES VIA RECIPROCAL LEARNING

This work tackles the problem of removing shadows from images using only high-level image labels, without theneed for detailed pixel-level annotations. The proposed method relies on a deep learning framework that learns todistinguish and eliminate shadows based on whether an image contains shadows or not. A key feature of thisapproach is the interactive optimization between two components—a shadow detector and a shadow remover—that work together to gradually enhance the model’s performance. Instead of depending on traditional supervisedmethods, a self-guided learning strategy is employed to ensure robust training without overfitting. Additionally,a specialized discriminator is introduced to help the model preserve color consistency and maintain realistic visualquality after shadow removal. Extensive evaluations on various datasets, including paired and unpaired samples,demonstrate the effectiveness and adaptability of this mutual learning-based framework.