Real noise image adjustment networks for saliency-aware stylistic color retouch

Automatic Image Adjustment (AIA) mainly aims to realize stylistic color retouch in images. Recent years have witnessed unprecedented success in learning-based AIA methods, especially convolutional neural networks (CNNs). However, existing AIA methods usually handle images without real noise from ideal scenarios, resulting in poor retouch performance when processing real noise images. Furthermore, these AIA methods lack attentive capability when learning salient areas to perform stylistic color retouch as human artists do. To address these problems, we first remodel the adjustment task for real noise images to remove the real noise. Then, we further propose the Real Noise Image Adjustment Networks (RNIA-Nets) using saliency-aware stylistic color retouch and adaptive denoising methods. Specifically, the saliency-aware stylistic color retouch predicts visual salient areas to learn stylistic color mapping using a proposed multifaceted attention (MFA) module. The adaptive denoising mechanism effectively predicts the denoising kernel for various real noise images. Eventually, to equitably verify the effectiveness of the proposed RNIA-Nets, a new challenging benchmark dataset collected from real noise images is established. Extensive experimental results demonstrate that the proposed method can achieve favorable results on real noise image adjustment, providing a highly effective solution to practical AIA applications. The code and datasets will be released at https://github.com/JiangBoCS/RNIA-Nets.