Edge Guided Low-Light Image Enhancement

Images captured under poorly lit environments often suffer from poor brightness, poor contrast, unwanted noise and color distortion. To enhance these images, there is a need to provide additional information while also reducing the burden of generalization on a single network. Therefore, this paper has proposed a novel fully convolutional two-step enhancement process consisting of EdgeNet and EnhanceNet. Edge-Net takes an under-exposed image as an input and predicts the edges in it's well-exposed image. The output from Edge-Net coupled with the low-light image is provided as input to EnhanceNet. EnhanceNet incorporates repeated use of Enhancement Blocks to eliminate noise, extract features from the low-light images and enhance these features to produce an output image with good visual characters. A major problem for low-light image enhancement is unavailability of paired data. To overcome this limitation, realistic low-light images are synthetically generated from well-exposed images. Extensive experiments show the efficacy of our method over the existing ones in qualitative as well as quantitative ways.