This work builds the connection between the regularity theory of optimal transportation map, Monge-Amp\`{e}re equation and GANs, which gives a theoretic understanding of the major drawbacks of GANs: convergence difficulty and mode collapse. According to the regularity theory of Monge-Amp\`{e}re equation, if the support of the target measure is disconnected or just non-convex, the optimal transportation mapping is discontinuous. General DNNs can only approximate continuous mappings. This intrinsic conflict leads to the convergence difficulty and mode collapse in GANs. We test our hypothesis that the supports of real data distribution are in general non-convex, therefore the discontinuity is unavoidable using an Autoencoder combined with discrete optimal transportation map (AE-OT framework) on the CelebA data set. The testing result is positive. Furthermore, we propose to approximate the continuous Brenier potential directly based on discrete Brenier theory to tackle mode collapse. Comparing with existing method, this method is more accurate and effective.
Underwater image enhancement is an important low-level vision task with many applications, and numerous algorithms have been proposed in recent years. Despite the demonstrated success, these results are often generated based on different assumptions using different datasets and metrics. In this paper, we propose a large-scale Realistic Underwater Image Enhancement (RUIE) dataset, in which all degraded images are divided into multiple sub-datasets according to natural underwater image quality evaluation metric and the degree of color deviation. Compared with exiting testing or training sets of realistic underwater scenes, the RUIE dataset contains three sub-datasets, which are specifically selected and classified for the experiment of non-reference image quality evaluation, color deviation and task-driven detection. Based on RUIE, we conduct extensive and systematic experiments to evaluate the effectiveness and limitations of various algorithms, on images with hierarchical classification of degradation. Our evaluation and analysis demonstrate the performance and limitations of state-of-the-art algorithms. The findings from these experiments not only confirm what is commonly believed, but also suggest new research directions. More importantly, we recognize that underwater image enhancement in practice usually serves as the preprocessing step for mid-level and high-level vision tasks. We thus propose to exploit the object detection performance on the enhanced images as a brand-new `task-specific' evaluation criterion for underwater image enhancement algorithms.
Magnetic Resonance Imaging (MRI) is one of the most dynamic and safe imaging techniques available for clinical applications. However, the rather slow speed of MRI acquisitions limits the patient throughput and potential indi cations. Compressive Sensing (CS) has proven to be an efficient technique for accelerating MRI acquisition. The most widely used CS-MRI model, founded on the premise of reconstructing an image from an incompletely filled k-space, leads to an ill-posed inverse problem. In the past years, lots of efforts have been made to efficiently optimize the CS-MRI model. Inspired by deep learning techniques, some preliminary works have tried to incorporate deep architectures into CS-MRI process. Unfortunately, the convergence issues (due to the experience-based networks) and the robustness (i.e., lack real-world noise modeling) of these deeply trained optimization methods are still missing. In this work, we develop a new paradigm to integrate designed numerical solvers and the data-driven architectures for CS-MRI. By introducing an optimal condition checking mechanism, we can successfully prove the convergence of our established deep CS-MRI optimization scheme. Furthermore, we explicitly formulate the Rician noise distributions within our framework and obtain an extended CS-MRI network to handle the real-world nosies in the MRI process. Extensive experimental results verify that the proposed paradigm outperforms the existing state-of-the-art techniques both in reconstruction accuracy and efficiency as well as robustness to noises in real scene.
Numerous tasks at the core of statistics, learning and vision areas are specific cases of ill-posed inverse problems. Recently, learning-based (e.g., deep) iterative methods have been empirically shown to be useful for these problems. Nevertheless, integrating learnable structures into iterations is still a laborious process, which can only be guided by intuitions or empirical insights. Moreover, there is a lack of rigorous analysis about the convergence behaviors of these reimplemented iterations, and thus the significance of such methods is a little bit vague. This paper moves beyond these limits and proposes Flexible Iterative Modularization Algorithm (FIMA), a generic and provable paradigm for nonconvex inverse problems. Our theoretical analysis reveals that FIMA allows us to generate globally convergent trajectories for learning-based iterative methods. Meanwhile, the devised scheduling policies on flexible modules should also be beneficial for classical numerical methods in the nonconvex scenario. Extensive experiments on real applications verify the superiority of FIMA.
Scribble colors based line art colorization is a challenging computer vision problem since neither greyscale values nor semantic information is presented in line arts, and the lack of authentic illustration-line art training pairs also increases difficulty of model generalization. Recently, several Generative Adversarial Nets (GANs) based methods have achieved great success. They can generate colorized illustrations conditioned on given line art and color hints. However, these methods fail to capture the authentic illustration distributions and are hence perceptually unsatisfying in the sense that they often lack accurate shading. To address these challenges, we propose a novel deep conditional adversarial architecture for scribble based anime line art colorization. Specifically, we integrate the conditional framework with WGAN-GP criteria as well as the perceptual loss to enable us to robustly train a deep network that makes the synthesized images more natural and real. We also introduce a local features network that is independent of synthetic data. With GANs conditioned on features from such network, we notably increase the generalization capability over "in the wild" line arts. Furthermore, we collect two datasets that provide high-quality colorful illustrations and authentic line arts for training and benchmarking. With the proposed model trained on our illustration dataset, we demonstrate that images synthesized by the presented approach are considerably more realistic and precise than alternative approaches.
Single image dehazing is an important low-level vision task with many applications. Early researches have investigated different kinds of visual priors to address this problem. However, they may fail when their assumptions are not valid on specific images. Recent deep networks also achieve relatively good performance in this task. But unfortunately, due to the disappreciation of rich physical rules in hazes, large amounts of data are required for their training. More importantly, they may still fail when there exist completely different haze distributions in testing images. By considering the collaborations of these two perspectives, this paper designs a novel residual architecture to aggregate both prior (i.e., domain knowledge) and data (i.e., haze distribution) information to propagate transmissions for scene radiance estimation. We further present a variational energy based perspective to investigate the intrinsic propagation behavior of our aggregated deep model. In this way, we actually bridge the gap between prior driven models and data driven networks and leverage advantages but avoid limitations of previous dehazing approaches. A lightweight learning framework is proposed to train our propagation network. Finally, by introducing a taskaware image separation formulation with a flexible optimization scheme, we extend the proposed model for more challenging vision tasks, such as underwater image enhancement and single image rain removal. Experiments on both synthetic and realworld images demonstrate the effectiveness and efficiency of the proposed framework.
Blind image deblurring plays a very important role in many vision and multimedia applications. Most existing works tend to introduce complex priors to estimate the sharp image structures for blur kernel estimation. However, it has been verified that directly optimizing these models is challenging and easy to fall into degenerate solutions. Although several experience-based heuristic inference strategies, including trained networks and designed iterations, have been developed, it is still hard to obtain theoretically guaranteed accurate solutions. In this work, a collaborative learning framework is established to address the above issues. Specifically, we first design two modules, named Generator and Corrector, to extract the intrinsic image structures from the data-driven and knowledge-based perspectives, respectively. By introducing a collaborative methodology to cascade these modules, we can strictly prove the convergence of our image propagations to a deblurring-related optimal solution. As a nontrivial byproduct, we also apply the proposed method to address other related tasks, such as image interpolation and edge-preserved smoothing. Plenty of experiments demonstrate that our method can outperform the state-of-the-art approaches on both synthetic and real datasets.
Currently, most top-performing text detection networks tend to employ fixed-size anchor boxes to guide the search for text instances. They usually rely on a large amount of anchors with different scales to discover texts in scene images, thus leading to high computational cost. In this paper, we propose an end-to-end box-based text detector with scale-adaptive anchors, which can dynamically adjust the scales of anchors according to the sizes of underlying texts by introducing an additional scale regression layer. The proposed scale-adaptive anchors allow us to use a few number of anchors to handle multi-scale texts and therefore significantly improve the computational efficiency. Moreover, compared to discrete scales used in previous methods, the learned continuous scales are more reliable, especially for small texts detection. Additionally, we propose Anchor convolution to better exploit necessary feature information by dynamically adjusting the sizes of receptive fields according to the learned scales. Extensive experiments demonstrate that the proposed detector is fast, taking only $0.28$ second per image, while outperforming most state-of-the-art methods in accuracy.
Deep learning is the mainstream technique for many machine learning tasks, including image recognition, machine translation, speech recognition, and so on. It has outperformed conventional methods in various fields and achieved great successes. Unfortunately, the understanding on how it works remains unclear. It has the central importance to lay down the theoretic foundation for deep learning. In this work, we give a geometric view to understand deep learning: we show that the fundamental principle attributing to the success is the manifold structure in data, namely natural high dimensional data concentrates close to a low-dimensional manifold, deep learning learns the manifold and the probability distribution on it. We further introduce the concepts of rectified linear complexity for deep neural network measuring its learning capability, rectified linear complexity of an embedding manifold describing the difficulty to be learned. Then we show for any deep neural network with fixed architecture, there exists a manifold that cannot be learned by the network. Finally, we propose to apply optimal mass transportation theory to control the probability distribution in the latent space.
Operator splitting methods have been successfully used in computational sciences, statistics, learning and vision areas to reduce complex problems into a series of simpler subproblems. However, prevalent splitting schemes are mostly established only based on the mathematical properties of some general optimization models. So it is a laborious process and often requires many iterations of ideation and validation to obtain practical and task-specific optimal solutions, especially for nonconvex problems in real-world scenarios. To break through the above limits, we introduce a new algorithmic framework, called Learnable Bregman Splitting (LBS), to perform deep-architecture-based operator splitting for nonconvex optimization based on specific task model. Thanks to the data-dependent (i.e., learnable) nature, our LBS can not only speed up the convergence, but also avoid unwanted trivial solutions for real-world tasks. Though with inexact deep iterations, we can still establish the global convergence and estimate the asymptotic convergence rate of LBS only by enforcing some fairly loose assumptions. Extensive experiments on different applications (e.g., image completion and deblurring) verify our theoretical results and show the superiority of LBS against existing methods.