Amplified Noise Map Guided Network for Low-Light Image Enhancement

Abstract

Low-light image is easily degraded by real noise, which brings great challenges for image enhancement task because the enhancement process will amplify the noise. To address this problem, we propose an amplified noise map guided network (AMG-Net), which simultaneously achieves the low-light enhancement and noise removal by extracting amplified noise map to guide the network training. Specifically, we build an encoder-decoder network as the basic enhancement model to get a preliminary enhanced image that usually includes amplified noise. Subsequently, we fed the preliminary enhanced image into a noise map estimator to continuously estimating the amplified noise map during the enhancement process by adopting residual connection. Finally, a residual block with adaptive instance normalization (AIN) is used to build a denoising model, which is guided by the noise map estimator to remove the amplified noise. Extensive experimental results demonstrate that the proposed AMG-Net can achieve competitive results compared with the existing state-of-the-art methods.