FPN-Retinex Enables Semi-Supervised Low-Light Image Enhancement Via Feature Pyramid Network with Flexible Backbones

Nowadays unsupervised low-light image enhancement methods reduce the demand for training data and the risk of overfitting. Yet they face challenges in achieving a balance between model performance and efficiency when handling real-world low-light images in unknown complex scenarios. Herein, we present a semi-supervised low-light image enhancement scheme termed FPN-Retinex that can leverage the supervision of a limited number of low-light/normal image pairs to realize an accurate Retinex decomposition, and based on this, achieve brightening the illumination of unpaired images to reduce dependence on paired datasets and improve generalization ability. The decomposition network is learned with some newly established constraints for complete decoupling between reflectance and illumination. For the first time, we introduce the feature pyramid network (FPN) to adjust the illumination maps of other low-light images without any supervision. Under this flexible framework, a wide range of backbones can be employed to work with illumination map generator, to navigate the balance between performance and efficiency. In addition, a novel attention mechanism is integrated into the FPN for giving the adaptability towards application scenes with different environment like underwater image enhancement (UIE) and dark face detection. Extensive experiments demonstrate that our proposed scheme has a more robust performance with high efficiency facing various images from different low-light environments over state-of-the-art methods.