FBGAN: multi-scale feature aggregation combined with boosting strategy for low-light image enhancement

Most of the existing low-light image enhancement methods focus only on enhancing the overall image brightness, ignoring the image details during the enhancement process, which leads to problems such as loss of image details and over-smoothing. In addition, the noise presented in the low-light image is still retained or even amplified after enhancement. This paper proposes a single-stage generative adversarial network, dubbed FBGAN, to address the above issues effectively. A multi-scale feature aggregation module based on an error feedback mechanism and a denoising module integrated with boosting strategy guided by attention mechanism are proposed in our model. The former preserves image details entirely during the enhancement, while the latter can simultaneously enhance low-light images and denoise. By these means, our model is competent to restore images with precise details, noise-free, distinct contrast and natural color. Extensive experiments are conducted to show the superiority of our model in terms of both qualitative and quantitative studies.