N2PN: Non-reference two-pathway network for low-light image enhancement

Many existing low-light image enhancement methods produce unnatural, low-contrast and color-distorted results. This paper presents a non-reference two-pathway network to solve the above problems. First, the frequency decomposition block is designed to decompose the input image into structure and detail pathways. Then, these two pathways are processed by different networks to improve the brightness and enhance the contrast. For the structure pathway, a light enhancement net is utilized to map the decomposed structure component to a set of adjustment parameters, which can adjust the range of the pixels. For the detail pathway, a contrast enhancement net is proposed to map the decomposed detail component into a series of parameters, which are iteratively applied in logarithmic image processing to obtain the final enhanced image. In these two pathways, the guide mechanism is adopted to maintain the color and spatial consistency of the enhanced result. In addition, the proposed method does not require any reference images during training, and it is a lightweight network that achieves low-light image enhancement. Extensive experiments demonstrate the advantages of the proposed method compared with state-of-the-art methods in both subjective and objective assessments.