Recovering images distorted by atmospheric turbulence is a challenging inverse problem due to the stochastic nature of turbulence. Although numerous turbulence mitigation (TM) algorithms have been proposed, their efficiency and generalization to real-world dynamic scenarios remain severely limited. Building upon the intuitions of classical TM algorithms, we present the Deep Atmospheric TUrbulence Mitigation network (DATUM). DATUM aims to overcome major challenges when transitioning from classical to deep learning approaches. By carefully integrating the merits of classical multi-frame TM methods into a deep network structure, we demonstrate that DATUM can efficiently perform long-range temporal aggregation using a recurrent fashion, while deformable attention and temporal-channel attention seamlessly facilitate pixel registration and lucky imaging. With additional supervision, tilt and blur degradation can be jointly mitigated. These inductive biases empower DATUM to significantly outperform existing methods while delivering a tenfold increase in processing speed. A large-scale training dataset, ATSyn, is presented as a co-invention to enable generalization in real turbulence. Our code and datasets will be available at \href{https://xg416.github.io/DATUM}{\textcolor{pink}{https://xg416.github.io/DATUM}}
Blind deconvolution problems are severely ill-posed because neither the underlying signal nor the forward operator are not known exactly. Conventionally, these problems are solved by alternating between estimation of the image and kernel while keeping the other fixed. In this paper, we show that this framework is flawed because of its tendency to get trapped in local minima and, instead, suggest the use of a kernel estimation strategy with a non-blind solver. This framework is employed by a diffusion method which is trained to sample the blur kernel from the conditional distribution with guidance from a pre-trained non-blind solver. The proposed diffusion method leads to state-of-the-art results on both synthetic and real blur datasets.
Non-blind image deconvolution has been studied for several decades but most of the existing work focuses on blur instead of noise. In photon-limited conditions, however, the excessive amount of shot noise makes traditional deconvolution algorithms fail. In searching for reasons why these methods fail, we present a systematic analysis of the Poisson non-blind deconvolution algorithms reported in the literature, covering both classical and deep learning methods. We compile a list of five "secrets" highlighting the do's and don'ts when designing algorithms. Based on this analysis, we build a proof-of-concept method by combining the five secrets. We find that the new method performs on par with some of the latest methods while outperforming some older ones.
At the pinnacle of computational imaging is the co-optimization of camera and algorithm. This, however, is not the only form of computational imaging. In problems such as imaging through adverse weather, the bigger challenge is how to accurately simulate the forward degradation process so that we can synthesize data to train reconstruction models and/or integrating the forward model as part of the reconstruction algorithm. This article introduces the concept of computational image formation (CIF). Compared to the standard inverse problems where the goal is to recover the latent image $\mathbf{x}$ from the observation $\mathbf{y} = \mathcal{G}(\mathbf{x})$, CIF shifts the focus to designing an approximate mapping $\mathcal{H}_{\theta}$ such that $\mathcal{H}_{\theta} \approx \mathcal{G}$ while giving a better image reconstruction result. The word ``computational'' highlights the fact that the image formation is now replaced by a numerical simulator. While matching nature remains an important goal, CIF pays even greater attention on strategically choosing an $\mathcal{H}_{\theta}$ so that the reconstruction performance is maximized. The goal of this article is to conceptualize the idea of CIF by elaborating on its meaning and implications. The first part of the article is a discussion on the four attributes of a CIF simulator: accurate enough to mimic $\mathcal{G}$, fast enough to be integrated as part of the reconstruction, providing a well-posed inverse problem when plugged into the reconstruction, and differentiable in the backpropagation sense. The second part of the article is a detailed case study based on imaging through atmospheric turbulence. The third part of the article is a collection of other examples that fall into the category of CIF. Finally, thoughts about the future direction and recommendations to the community are shared.
Image distortion by atmospheric turbulence is a stochastic degradation, which is a critical problem in long-range optical imaging systems. A number of research has been conducted during the past decades, including model-based and emerging deep-learning solutions with the help of synthetic data. Although fast and physics-grounded simulation tools have been introduced to help the deep-learning models adapt to real-world turbulence conditions recently, the training of such models only relies on the synthetic data and ground truth pairs. This paper proposes the Physics-integrated Restoration Network (PiRN) to bring the physics-based simulator directly into the training process to help the network to disentangle the stochasticity from the degradation and the underlying image. Furthermore, to overcome the ``average effect" introduced by deterministic models and the domain gap between the synthetic and real-world degradation, we further introduce PiRN with Stochastic Refinement (PiRN-SR) to boost its perceptual quality. Overall, our PiRN and PiRN-SR improve the generalization to real-world unknown turbulence conditions and provide a state-of-the-art restoration in both pixel-wise accuracy and perceptual quality. Our codes are available at \url{https://github.com/VITA-Group/PiRN}.
The advancement of new digital image sensors has enabled the design of exposure multiplexing schemes where a single image capture can have multiple exposures and conversion gains in an interlaced format, similar to that of a Bayer color filter array. In this paper, we ask the question of how to design such multiplexing schemes for adaptive high-dynamic range (HDR) imaging where the multiplexing scheme can be updated according to the scenes. We present two new findings. (i) We address the problem of design optimality. We show that given a multiplex pattern, the conventional optimality criteria based on the input/output-referred signal-to-noise ratio (SNR) of the independently measured pixels can lead to flawed decisions because it cannot encapsulate the location of the saturated pixels. We overcome the issue by proposing a new concept known as the spatially varying exposure risk (SVE-Risk) which is a pseudo-idealistic quantification of the amount of recoverable pixels. We present an efficient enumeration algorithm to select the optimal multiplex patterns. (ii) We report a design universality observation that the design of the multiplex pattern can be decoupled from the image reconstruction algorithm. This is a significant departure from the recent literature that the multiplex pattern should be jointly optimized with the reconstruction algorithm. Our finding suggests that in the context of exposure multiplexing, an end-to-end training may not be necessary.
While today's high dynamic range (HDR) image fusion algorithms are capable of blending multiple exposures, the acquisition is often controlled so that the dynamic range within one exposure is narrow. For HDR imaging in photon-limited situations, the dynamic range can be enormous and the noise within one exposure is spatially varying. Existing image denoising algorithms and HDR fusion algorithms both fail to handle this situation, leading to severe limitations in low-light HDR imaging. This paper presents two contributions. Firstly, we identify the source of the problem. We find that the issue is associated with the co-existence of (1) spatially varying signal-to-noise ratio, especially the excessive noise due to very dark regions, and (2) a wide luminance range within each exposure. We show that while the issue can be handled by a bank of denoisers, the complexity is high. Secondly, we propose a new method called the spatially varying high dynamic range (SV-HDR) fusion network to simultaneously denoise and fuse images. We introduce a new exposure-shared block within our custom-designed multi-scale transformer framework. In a variety of testing conditions, the performance of the proposed SV-HDR is better than the existing methods.
In this paper, we study how pretraining label granularity affects the generalization of deep neural networks in image classification tasks. We focus on the "fine-to-coarse" transfer learning setting where the pretraining label is more fine-grained than that of the target problem. We experiment with this method using the label hierarchy of iNaturalist 2021, and observe a 8.76% relative improvement of the error rate over the baseline. We find the following conditions are key for the improvement: 1) the pretraining dataset has a strong and meaningful label hierarchy, 2) its label function strongly aligns with that of the target task, and most importantly, 3) an appropriate level of pretraining label granularity is chosen. The importance of pretraining label granularity is further corroborated by our transfer learning experiments on ImageNet. Most notably, we show that pretraining at the leaf labels of ImageNet21k produces better transfer results on ImageNet1k than pretraining at other coarser granularity levels, which supports the common practice. Theoretically, through an analysis on a two-layer convolutional ReLU network, we prove that: 1) models trained on coarse-grained labels only respond strongly to the common or "easy-to-learn" features; 2) with the dataset satisfying the right conditions, fine-grained pretraining encourages the model to also learn rarer or "harder-to-learn" features well, thus improving the model's generalization.
A spatially varying blur kernel $h(\mathbf{x},\mathbf{u})$ is specified by an input coordinate $\mathbf{u} \in \mathbb{R}^2$ and an output coordinate $\mathbf{x} \in \mathbb{R}^2$. For computational efficiency, we sometimes write $h(\mathbf{x},\mathbf{u})$ as a linear combination of spatially invariant basis functions. The associated pixelwise coefficients, however, can be indexed by either the input coordinate or the output coordinate. While appearing subtle, the two indexing schemes will lead to two different forms of convolutions known as scattering and gathering, respectively. We discuss the origin of the operations. We discuss conditions under which the two operations are identical. We show that scattering is more suitable for simulating how light propagates and gathering is more suitable for image filtering such as denoising.