<p>High-quality images are essential for computer vision tasks. Low-light images often suffer from various degradations due to limited capture conditions, which adversely affect subsequent tasks. In this work, we propose a simple yet effective low-light image enhancement(LLIE) method, termed UALIE, which is inspired by the Retinex theory and decomposes the original image into illumination and reflectance components. First, to extract rich features at multiple levels, the network adopts a U-Net inspired design that fuses features from multiple branches to capture multi-level information. Second, to further guide the network toward more informative features, we introduce the attention mechanism that enables the model to automatically focus on important regions. Finally, our method does not require normal-light reference images during training, but instead learns from paired low-light instances only. This significantly reduces the dependency on extensive labeled data. Extensive experiments demonstrate that UALIE outperforms state-of-the-art techniques on several public datasets.</p>