Dual-band low-light image enhancement

Most of the existing low-light image enhancement algorithms are designed for one kind of low-light image, which cannot effectively handle the different exposed parts of an image. In this paper, we propose a novel dual-band low-light image enhancement algorithm that utilizes a multiple-constrained dual-band light enhancement curve to differentiate the different exposed parts of an image, enhance low light, maintain normal light, and suppress overexposure to achieve image enhancement. We enrich the illumination information of the training data set by the alternate preprocessing module and design a multi-constrained dual-band light enhancement curve for image enhancement based on the characteristics of various low-light images. Then, the enhancement curve is optimized again by guiding the deep learning network through non-reference gradient exposure loss. Non-reference gradient exposure loss evaluates the exposure loss of the enhanced image based on the judgment of the gradient difference of the input image. The image's brightness is converged to a reasonable range by continuously iterating the dual-band light enhancement curve. Experiments on various benchmarks show that our method outperforms other state-of-the-art algorithms.