Dual-Teacher Distillation for Low-Light Image Enhancement
Jeong-Hyeok Park, Tae-Hyeon Kim, Jong‐Ok Kim · 2022 Asia-Pacific Signal and Information Processing Association Annual Summit and Conference (APSIPA ASC) · 2022
Low-light enhancement studies based on the Retinex theory have been widely conducted. Previous studies show that enhancing low-light images through adjustment of the reflectance and illumination components is one of the promising approaches. In this paper, we extended the conventional Retinex-Net with dual-teacher distillation and attention mechanisms for further improvement. The attention mechanism is adopted in auto-encoder architecture to maximize enhancement performance. Also, we create our own dataset with different exposure times which is suitable for effectively training the proposed network. Experimental results show that the proposed dual-teacher knowledge distillation outperforms Retinex-Net and other existing methods, particularly for ultra low-light scenes.