Decomposing-Recomposing Network: A Novel Reconstruction and Enhancement Approach for Low-Light Image

Abstract

Low-light imaging has long been a fundamental yet challenging area in optical imaging, primarily due to photon-starved conditions that not only impair image visibility but also amplify sensor noise. Traditional enhancement techniques mainly focus on illumination adjustments and often fall short in addressing the inherent trade-off between boosting brightness and suppressing noise. Moreover, many existing methods assume expert-level noise management during image capture, thereby overlooking the crucial frequency-dependent characteristics of noise distribution. Our observations reveal that while simultaneously enhancing brightness and reducing noise is challenging, noise intensity indeed varies across different frequency layers. In low-light images, the noise predominantly appears as Additive White Gaussian Noise (AWGN) concentrated in the high-frequency domain. Inspired by this phenomenon, we introduce a novel model based on the principles of decomposition and recomposition. Extensive experiments conducted on several baseline low-light datasets—including LOL-v1, LOL-v2, MEF, LIME, DICM, and NPE—demonstrate that our approach not only outperforms the latest methods quantitatively and qualitatively but also excels in handling real and complex low-light scenarios. Furthermore, our method consistently produces superior visual outcomes compared to existing techniques.