Label-Noise Robust Generative Adversarial Networks

Abstract

Recently, generative adversarial networks (GANs), which learn data distributions through adversarial training, have gained special attention owing to their high image reproduction ability. However, one limitation of standard GANs is that they recreate training images faithfully despite image degradation characteristics such as blur, noise, and compression. To remedy this, we address the problem of blur, noise, and compression robust image generation. Our objective is to learn a non-degraded image generator directly from degraded images without prior knowledge of image degradation. The recently proposed noise robust GAN (NR-GAN) already provides a solution to the problem of noise degradation. Therefore, we first focus on blur and compression degradations. We propose blur robust GAN (BR-GAN) and compression robust GAN (CR-GAN), which learn a kernel generator and quality factor generator, respectively, with non-degraded image generators. Owing to the irreversible blur and compression characteristics, adjusting their strengths is non-trivial. Therefore, we incorporate switching architectures that can adapt the strengths in a data-driven manner. Based on BR-GAN, NR-GAN, and CR-GAN, we further propose blur, noise, and compression robust GAN (BNCR-GAN), which unifies these three models into a single model with additionally introduced adaptive consistency losses that suppress the uncertainty caused by the combination. We provide benchmark scores through large-scale comparative studies on CIFAR-10 and a generality analysis on FFHQ dataset.1 More precisely, to reduce this cost, unpaired learning methods [53,50] were devised.These methods, however, still require separate collection of non-degraded and degraded images. As an alternative, self-supervised learning methods [45, 2, 48] were also proposed; however, their application is limited to denoising.Blur, Noise, and Compression Robust Generative Adversarial Networks 3 Motivated by these existing studies, we address the following problem: "How can we learn a non-degraded image generator directly from degraded images without knowing the details of image degradation? " Particularly, in this work, we focus on three representative types of image degradation, i.e., blur, noise, and compression, and call this problem blur, noise, and compression robust image generation. As an example, we demonstrate a solution using our model in Fig. 1(c). Although the same training images ( Fig. 1(a)) are used between baseline GAN (Fig. 1(b)) and proposed blur, noise, and compression robust GAN (BNCR-GAN ) (Fig. 1(c)), BNCR-GAN succeeds in generating non-degraded images, whereas baseline GAN fails to do so.The recently proposed noise robust GAN (NR-GAN) [31] has shown promising results when its application is restricted to noise degradation. Hence, in this work, we first focus on the remaining two kinds of degradation, i.e., blur and compression, and handle all types of degradation as an advanced problem.To obtain robustness in blur and compression degradations, we first propose two variants of ...

EnerGAN++: A Generative Adversarial Gated Recurrent Network for Robust Energy Disaggregation

“…2) Generative Adversarial Networks: GANs are popular in a variety of application domains, including photorealistic image super-resolution [23], image inpainting [24], text to image synthesis [25]. Several studies have shown promising results for reproducing data, in spite of labels corrupted by random noise [7], [26], [27]. Generative adversarial networks learn a deep generative model that is able to synthesize high dimensional data samples.…”

“…Generative adversarial networks (GAN) [6], which is a special category of generative models, can learn data distributions through adversarial learning, even in the presence of noise. Therefore, they are capable of representing noisy time series with high consistency [7]. Given that GAN input is a noise vector, the model is inherently more robust to noisy input signals and less sensitive to inconsistencies or gaps in the training input samples.…”

Correcting and reweighting false label masks in brain tumor segmentation

“…6 However, since the training data are selected on the fly rather than in the beginning, it is hard to characterize these sample-selection biases, and it is also difficult to give any theoretical guarantee on the learning consistency. The noise transition matrix-based approaches 5 assume that there is a single transition probability between the noisy label and the ground-truth label, and this probability is independent of the individual samples. However, in real-world cases, the appearance of each sample has much influence on whether it will be misclassified.…”

“…Note that the top row illustrates that the tumor core (TC) is easily annotated as the whole tumor (WT), and the bottom row shows that the enhancing tumor (ET) has a high similarity with the TC. Most approaches 5 model the probability of the label masks with errors and learn the error transition matrix (the probability of the label masks with errors conditioned the true label mask). However, in real cases, the appearance of each sample has much influence on whether it can be misclassified.…”

“…Most approaches 5 model the probability of the label masks with errors and learn the error transition matrix (the probability of the label masks with errors conditioned the true label mask). However, in real cases, the appearance of each sample has much influence on whether it can be misclassified.…”