In this paper, we propose a structure-based low-rank Retinex model for simultaneous low-light image enhancement and noise removal. Based on the traditional variational-based Retinex framework, in the proposed model, a smooth prior is forced on the illumination, and a gradient fidelity term and the weighted nuclear norm are used to suppress noise and enhance structural details in the reflectance. By considering that the manifold structure similarity is more effective than intensity similarity in describing the structural features of image patches, we further propose to use the manifold structure similarity in image patch grouping. Then, an alternating direction minimization algorithm is used to solve the reflectance estimating model. The entire process for solving the proposed model uses a sequential optimization. The final enhancement results is obtained by combining the reflectance and the Gamma corrected illumination. Experiment show that, the proposed method can simultaneously enhance and denoise the low-light image, and produce better or comparable results compared with the state-of-the-art methods
This paper proposes a deep convolutional neural network-based low-light image enhancement method. In order to adaptively enhance the image brightness, a convolutional neural network with convolutional modules is designed. Lowlight image is firstly down-sampled into sub-images. Then an illumination map is obtained from the input image to provide additional information to the network. The network works on a tensor that consists of sub-images and illumination map, achieving a good performance in brightness increasing and structure preservation. The enhanced result is reconstructed from the output sub-images. Experimental results demonstrate the effectiveness of the proposed method in low-light image enhancement.
In the traditional uniform blind deblurring methods, we have witnessed the great advances by utilizing various image priors which are expected to favor clean images than blurred images and act through regularizing the solution space. However, these methods failed in dealing with non-uniform blind deblurring because of the inaccuracy in kernel estimation. Learning-based methods can generate clear images in an end-to-end way potentially without an intermediate step of blur kernel estimation. To better deal with the non-uniform deblurring problem in dynamic scenes, in this paper we present a new type of image priors complementary to the deep learning-based blind estimation framework. Specifically, inspired by the interesting discovery of dark and bright channels in dehazing, the opposite-channel-based discriminative priors are developed and directly integrated to the loss of our advocated deep deblurring model, so as to achieve more accurate and robust blind deblurring performance. It deserves noticing that, our deep model is formulated in the framework of the Wasserstein generative adversarial networks regularized by the Liptchitz penalty (WGAN-LP), and the network structures are relatively simpler yet more stable than other deep deblurring methods. We evaluate the proposed method on a large scale blur dataset with complex non-uniform motions. Experimental results show that it achieves state-of-the-art non-uniform blind deblurring performance not only quantitatively but also qualitatively
Persons captured in real-life scenarios are generally in non-uniform scales. However, most generally acknowledged person re-identification (Re-ID) methods lay emphasis on matching normal-scale high-resolution person images. To address this problem, the ideas of existing image reconstruction techniques are incorporated which are expected contribute to recover accurate appearance information for low-resolution person Re-ID. In specific, this paper proposes a joint deep learning approach for Scale-Adaptive person Super-Resolution and Re-identification (SASR2 ). It is for the first time that scale-adaptive learning is jointly implemented for super-resolution and re-identification without any extra post-processing process. With the super-resolution module, the high-resolution appearance information can be automatically reconstructed from scales of low-resolution person images, bringing a direct beneficial impact on the subsequent Re-ID thanks to the joint learning nature of the proposed approach. It deserves noting that SASR2 is not only simple but also flexible, since it can be adaptable to person Re-ID on both multi-scale LR and normal-scale HR datasets. A large amount of experimental analysis demonstrates that SASR2 achieves competitive performance compared with previous low-resolution Re-ID methods especially on the realistic CAVIAR dataset.
Proc. SPIE. 11069, Tenth International Conference on Graphics and Image Processing (ICGIP 2018)
KEYWORDS: Super resolution, Visualization, Image segmentation, Matrices, Image processing, Image quality, Telecommunications, Algorithm development, Communication engineering, Current controlled current source
This paper proposes a new variational model for deblurring low-resolution images, a.k.a. single image nonparametric blind super-resolution. In specific, a type of new adaptive heavy-tailed image priors are presented incorporating both the model discriminativeness and effectiveness of salient edge pursuit for accurate and reliable blur kernel estimation. With the assistance of appropriate non-blind super-resolution approaches, nonparametric blind super-resolution can be cast as a regularized functional minimization problem. An efficient numerical algorithm is derived by harnessing the alternating direction method of multipliers as well as the conjugate gradient method, with which alternatingly iterative estimations for kernel and image are finally implemented in a multi-scale manner. Numerous experiments are conducted along with comparisons made among the proposed approach and two recent state-of-the-art ones, demonstrating that the proposed approach is able to better deal with low-resolution images which are blurred by various possible kernels, e.g., Gaussianshaped kernels of varying sizes, ellipse-shaped kernels of varying orientations, curvilinear kernels of varying trajectories.
It is known that actual performance of most previous face hallucination approaches will drop dramatically as a very low-resolution tiny face is provided. Inspired by the latest progress in deep unsupervised learning, this paper works on tiny faces of size 16×16 pixels and magnifies them into their 8× upsampling ones by exploiting the boundary equilibrium generative adverarial networks (BEGAN). Besides imposing a pixel-wise L2 regularization term to the generative model, it is found that our targeted auto-encoding generator with residual blocks and skip connections is a key component for BEGAN achieving state-of-the-art hallucination performance. The cropped CelebA face dataset is preliminarily used in our experiments. The results demonstrate that the proposed approach is not only of fast and stable convergence, but also robust to pose, expression, illuminance and occluded variations.