Imaging in low-light environments has strong noise and low contrast, which can easily be exacerbated during image enhancement. RAW images have advantages over SRGB in low-light enhancement due to their linearity and other advantages, but they have the problem of image misalignment in the mapping to SRGB, and image features are lost with the model stream. In order to overcome these problems, we propose a supervised learning approach, which is a parallel two-stage network PTS-NET with RAW supervision and SRGB supervision. In the color restoration stage, we design a conversion module (FGM) that can extract high-frequency low-frequency feature information to achieve a smooth conversion from RAW to sRGB. Extensive experiments on three benchmark datasets demonstrate that our method not only achieves state-of-the-art performance but also enhances the quality of low-light images in practical applications, such as nighttime object recognition.