DB-GAN: A Low Contrast Image Enhancer Based on NIR-RGB Fusion
Linruize Tang, Longbin Yan, Jie Chen · 2022
RGB images captured under haze or over-/under-exposure conditions frequently have low contrast and lack of detail. Due to the limited information in the original image, the majority of enhancement techniques that rely solely on visible information fail to restore the original image satisfactorily. This emphasizes the need for information beyond the visible spectrum. In this paper, we formulate the low contrast image enhancement problem based on near-infrared (NIR)-RGB fusion. A Dual-Branch Generative Adversarial Network (DB-GAN) is designed based on the specific characteristics of NIR-RGB fusion problem. To be specific, with the guidance of the two discriminators that respectively extract information from RGB and NIR images, a U-net based generator generates informative, high-quality fused images. In addition, we create an NIR-RGB dataset with over 1300 aligned image pairs for training the network. Quantitative and qualitative experimental results show the superior performance of our proposed framework.