Adversarial Training for Dual-Stage Image Denoising Enhanced with Feature Matching

摘要

We propose a dual-stage convolutional neural network, augmented with adversarial training, to address the shortcoming of current convolutional neural networks in image denoising. Our dual-stage approach, coupled with feature matching, is especially effective in recovering fine detail under high noise level. First, we use residual learning denoising to output a preliminary denoised reference image. Then, an image reconstruction denoiser uses a multi-scale feature selection layer, which deploys skip-connections and ResNet blocks to recover the image detail based on the noisy image and the reference image. This dual-stage denoising is augmented with the feedback from a discriminator, which forms an adversarial training framework and guides the denoising towards a clean image construction. The feature matching process embedded in the discriminator ensures that the framework can be generalized to a diverse collection of image content. Experimental results show better denoising performance in public benchmark datasets compared with the state-of-the-art approaches.