Image denoising technology removes noise from the corrupted image by utilizing different features between image and noise. Convolutional neural network (CNN)-based algorithms have been the concern of the recent progress on diverse image restoration problems and become an efficient solution in image denoising. Although a quite number of existing CNN-based image denoising methods perform well on the simplified additive white Gaussian noise (AWGN) model, their performance often degrades severely on the real-world noisy images which are corrupted by more complicated noise. In this paper, we utilized the multi-task learning (MTL) framework to integrate multiple loss functions for collaborative training of CNN. This approach aims to improve the denoising performance of CNNs on real-world images with non-Gaussian noise. Simultaneously, to automatically optimize the weights of individual sub-tasks within the MTL framework, we incorporated a self-learning weight layer into the CNN. Extensive experiments demonstrate that our approach effectively enhances the denoising performance of CNN-based image denoising algorithms on real-world images. It reduces excessive image smoothing, improves quantitative metrics, and enhances visual quality in the restored images. Our method shows the effectiveness of the improved performance of denoising CNNS for real-world image denoising processing.
Generative adversarial network (GAN) has been widely applied to produce super-resolution (SR) image with real perception and texture details. In most existing SR approaches, the training objective typically measures a pixel-wise average distance between the SR and high-resolution (HR) images. However, as the degradation function of different images from HR to low resolution (LR) is generally different, optimizing such metrics often leads to certain unpleasant artificial traces. Unlike the prevalent GAN inversion methods that require expensive image-specific optimization at runtime, we present an alternative formulation by directly leveraging latent representation produced by a pretrained AutoEncoder. We call this improved method reduce dimension super-resolved GAN (RD-SRGAN). RD-SRGAN first obtains the latent feature representation of LR image by a pretrained AutoEncoder as input to the generator network. This process not only reduces noise effects but also decreases the overall computational complexity. On the other hand, the residual between the ground truth and the produced images replaces the produced images as input to the discriminator network, and a 2D zero mean Gaussian noise with controllable low variance replaces the real images as another input to the discriminator network. By leveraging the feature representation and properties of the 2D zero mean Gaussian noise, we restrict the optimization space to produce an SR image. Therefore, the residual of the generated SR images tend to approximates to a Gaussian noise, which introduces useless deviation information as little as possible. Experimental results show that RD-SRGAN can benefit from these strategies and achieve improved fidelity and naturalness comparison to existing methods. Switching the pretrained AutoEncoder allows the method to deal with images from diverse categories, e.g., remote sensing satellite imaging, medical imaging, and astronomy.