Deep Combined Image Denoising with Cloud Images.

Image denoising methods essentially lose some high-frequency (HF) information in denoising. To address this issue, we propose an end-to-end trainable deep network to additionally utilize online retrieved cloud images to compensate for the HF information loss based on the internal inferred results. In particular, the noise inference network first infers a noise map from the noisy image and derives an intermediate image by removing the noise map from the noisy image. Then the external online compensation is performed based on the intermediate image. The final results are obtained by fusing the intermediate image with external HF maps extracted by the external HF compensation network. Extensive experimental results demonstrate that our method achieves notably better performance than state-of-the-art denoising methods.