Deep neural networks (DNNs) have achieved significant success in image restoration tasks by directly learning a powerful non-linear mapping from corrupted images to their latent clean ones. However, there still exist two major limitations for these deep learning (DL)-based methods. Firstly, the noises contained in real corrupted images are very complex, usually neglected and largely under-estimated in most current methods. Secondly, existing DL methods are mostly trained on one pre-assumed degradation process for all of the training image pairs, such as the widely used bicubic downsampling assumption in the image super-resolution task, inevitably leading to poor generalization performance when the true degradation does not match with such assumed one. To address these issues, we propose a unified generative model for the image restoration, which elaborately configures the degradation process from the latent clean image to the observed corrupted one. Specifically, different from most of current methods, the pixel-wisely non-i.i.d. Gaussian distribution, being with more flexibility, is adopted in our method to fit the complex real noises. Furthermore, the method is built on the general image degradation process, making it capable of adapting diverse degradations under one single model. Besides, we design a variational inference algorithm to learn all parameters involved in the proposed model with explicit form of objective loss. Specifically, beyond traditional variational methodology, two DNNs are employed to parameterize the posteriori distributions, one to infer the distribution of the latent clean image, and another to infer the distribution of the image noise. Extensive experiments demonstrate the superiority of the proposed method on three classical image restoration tasks, including image denoising, image super-resolution and JPEG image deblocking.
Single image deraining is an important yet challenging issue due to the complex and diverse rain structures in real scenes. Currently, the state-of-the-art performance on this task is achieved by deep learning (DL)-based methods that mainly benefit from abundant pre-collected paired rainy-clean samples either manually synthesized or semi-automatically generated under human supervision. This tends to bring a large labor for data collection and more importantly, such manner neglects to elaborately explore the intrinsic generative mechanism of rain streaks which should be related to the most insightful understanding of the task. Against this issue, we investigate the generative process of rainy image and construct a full Bayesian generative model for generating rains from automatically extracted latent variables that represent physical structural factors for depicting rains, like direction, scale, and thickness. To solve this model, we propose an algorithm where the posteriors of latent variables are parameterized as CNNs and all the involved parameters can be inferred under a concise variational inference framework in a data-driven manner. Especially, the rain layer is modeled as an implicit distribution, parameterized as a generator, which avoids subjective prior assumptions on rains as in traditional model-based methods. More practically, from the learned generator, rain patches can be automatically generated and utilized to simulate diverse training pairs so as to enrich and augment the existing benchmark datasets. Comprehensive experiments substantiate that the proposed model has fine capability of generating plausible samples that not only helps significantly improve the deraining performance of current DL-based single image derainers, but also largely loosens the requirement of large training sample pre-collection for the task.
Deep neural networks often degrade significantly when training data suffer from class imbalance problems. Existing approaches, e.g., re-sampling and re-weighting, commonly address this issue by rearranging the label distribution of training data to train the networks fitting well to the implicit balanced label distribution. However, most of them hinder the representative ability of learned features due to insufficient use of intra/inter-sample information of training data. To address this issue, we propose meta feature modulator (MFM), a meta-learning framework to model the difference between the long-tailed training data and the balanced meta data from the perspective of representation learning. Concretely, we employ learnable hyper-parameters (dubbed modulation parameters) to adaptively scale and shift the intermediate features of classification networks, and the modulation parameters are optimized together with the classification network parameters guided by a small amount of balanced meta data. We further design a modulator network to guide the generation of the modulation parameters, and such a meta-learner can be readily adapted to train the classification network on other long-tailed datasets. Extensive experiments on benchmark vision datasets substantiate the superiority of our approach on long-tailed recognition tasks beyond other state-of-the-art methods.
Recent deep neural networks (DNNs) can easily overfit to biased training data with noisy labels. Label correction strategy is commonly used to alleviate this issue by designing a method to identity suspected noisy labels and then correct them. Current approaches to correcting corrupted labels usually need certain pre-defined label correction rules or manually preset hyper-parameters. These fixed settings make it hard to apply in practice since the accurate label correction usually related with the concrete problem, training data and the temporal information hidden in dynamic iterations of training process. To address this issue, we propose a meta-learning model which could estimate soft labels through meta-gradient descent step under the guidance of noise-free meta data. By viewing the label correction procedure as a meta-process and using a meta-learner to automatically correct labels, we could adaptively obtain rectified soft labels iteratively according to current training problems without manually preset hyper-parameters. Besides, our method is model-agnostic and we can combine it with any other existing model with ease. Comprehensive experiments substantiate the superiority of our method in both synthetic and real-world problems with noisy labels compared with current SOTA label correction strategies.
The learning rate (LR) is one of the most important hyper-parameters in stochastic gradient descent (SGD) for deep neural networks (DNNs) training and generalization. However, current hand-designed LR schedules need to manually pre-specify schedule as well as its extra hyper-parameters, which limits its ability to adapt non-convex optimization problems due to the significant variation of training dynamic. To address this issue, we propose a model capable of adaptively learning LR schedule from data. We specifically design a meta-learner with explicit mapping formulation to parameterize LR schedules, which can adjust LR adaptively to comply with current training dynamic by leveraging the information from past training histories. Image and text classification benchmark experiments substantiate the capability of our method for achieving proper LR schedules compared with baseline methods. Moreover, we transfer the learned LR schedule to other various tasks, like different training batch sizes, epochs, datasets, network architectures, especially large scale ImageNet dataset, showing its stronger generalization capability than related methods. Finally, guided by a small set of clean validation set, we show our method can achieve better generalization error when training data is biased with corrupted noise than baseline methods.
The recent advancement of deep learning techniques has made great progress on hyperspectral image super-resolution (HSI-SR). Yet the development of unsupervised deep networks remains challenging for this task. To this end, we propose a novel coupled unmixing network with a cross-attention mechanism, CUCaNet for short, to enhance the spatial resolution of HSI by means of higher-spatial-resolution multispectral image (MSI). Inspired by coupled spectral unmixing, a two-stream convolutional autoencoder framework is taken as backbone to jointly decompose MS and HS data into a spectrally meaningful basis and corresponding coefficients. CUCaNet is capable of adaptively learning spectral and spatial response functions from HS-MS correspondences by enforcing reasonable consistency assumptions on the networks. Moreover, a cross-attention module is devised to yield more effective spatial-spectral information transfer in networks. Extensive experiments are conducted on three widely-used HS-MS datasets in comparison with state-of-the-art HSI-SR models, demonstrating the superiority of the CUCaNet in the HSI-SR application. Furthermore, the codes and datasets will be available at: https://github.com/danfenghong/ECCV2020_CUCaNet.
Real-world image noise removal is a long-standing yet very challenging task in computer vision. The success of deep neural network in denoising stimulates the research of noise generation, aiming at synthesizing more clean-noisy image pairs to facilitate the training of deep denoisers. In this work, we propose a novel unified framework to simultaneously deal with the noise removal and noise generation tasks. Instead of only inferring the posteriori distribution of the latent clean image conditioned on the observed noisy image in traditional MAP framework, our proposed method learns the joint distribution of the clean-noisy image pairs. Specifically, we approximate the joint distribution with two different factorized forms, which can be formulated as a denoiser mapping the noisy image to the clean one and a generator mapping the clean image to the noisy one. The learned joint distribution implicitly contains all the information between the noisy and clean images, avoiding the necessity of manually designing the image priors and noise assumptions as traditional. Besides, the performance of our denoiser can be further improved by augmenting the original training dataset with the learned generator. Moreover, we propose two metrics to assess the quality of the generated noisy image, for which, to the best of our knowledge, such metrics are firstly proposed along this research line. Extensive experiments have been conducted to demonstrate the superiority of our method over the state-of-the-arts both in the real noise removal and generation tasks. The training and testing code is available at https://github.com/zsyOAOA/DANet.
To discover intrinsic inter-class transition probabilities underlying data, learning with noise transition has become an important approach for robust deep learning on corrupted labels. Prior methods attempt to achieve such transition knowledge by pre-assuming strongly confident anchor points with 1-probability belonging to a specific class, generally infeasible in practice, or directly jointly estimating the transition matrix and learning the classifier from the noisy samples, always leading to inaccurate estimation misguided by wrong annotation information especially in large noise cases. To alleviate these issues, this study proposes a new meta-transition-learning strategy for the task. Specifically, through the sound guidance of a small set of meta data with clean labels, the noise transition matrix and the classifier parameters can be mutually ameliorated to avoid being trapped by noisy training samples, and without need of any anchor point assumptions. Besides, we prove our method is with statistical consistency guarantee on correctly estimating the desired transition matrix. Extensive synthetic and real experiments validate that our method can more accurately extract the transition matrix, naturally following its more robust performance than prior arts. Its essential relationship with label distribution learning is also discussed, which explains its fine performance even under no-noise scenarios.