Enhancing Compressed Low-Light Image via Decoupling Strategy and Robust Priors

Low-Light Image Enhancement (LLIE) is crucial for enhancing the visual quality of images captured in dim lighting conditions, particularly considering the widespread mobile photography. Meanwhile, image compression is an essential component in mobile photography to balance image quality and storage space. However, most existing LLIE methods overlook the impact of compression, leading to significant performance degradation. Therefore, we draw the first attention to Compressed Low-Light Image Enhancement (CLLIE) and propose our solution. In contrast to many existing methods employing end-to-end architectures for direct learning of complex mappings, we introduce a two-decoder decoupling strategy aimed at achieving efficient decoupling and enhanced mapping learning. To mitigate performance degradation caused by the insufficient robustness of existing priors, we integrate two robust priors with our proposed Brightness-Guided Transformer (BGT) and Semantic-Guided Channel Attention Block (SGCAB), facilitating robust enhancement. Extensive experiments demonstrate a significant improvement of our proposed method over existing SOTAs, achieving up to 1.8 dB higher PSNR, which indicates the effectiveness of our design in suppressing the effect of compression and highlights the importance of considering compression in LLIE.