Full length articleOptimization and regularization of complex task decomposition for blind removal of multi-factor degradation☆
Graphical abstract
Introduction
Restoration of degraded images is a low-level vision task. It can help sensors obtain high quality images and has wide applications, such as image deblurring [1], [2], [3], [4], [5], image denoising [6], [7], image dehazing [8], image rain removal [9], and image super-resolution [10], [11], [12].
For these low-level computer vision tasks, deep learning has achieved remarkable success [13]. However, these successes mainly depend on the continuous innovation of deep neural network structure [13]. Xu et al. proposed a deep convolutional neural network (CNN) to capture the characteristics of degradation [14], which can recognize different types of the blurred images, thereby achieving excellent restoration results under certain conditions. Zhang et al. used a learning gradient method to guide image deconvolution [15]. Hradis et al. restored highly structured text through CNNs and achieved competitive results in the blind restoration of text [16], Zhang et al. used an end-to-end recurrent neural network (RNN) network to learn the spatial distribution involved in motion blur [5], while Guo et al. blindly removed noise from degraded images by using an estimated noise map as the input [7]. Nah et al. developed a multi-scale CNN that can restore sharp images in an end-to-end manner when blur is caused by various sources [1]. Mao et al. achieved good results using a very deep skip-link CNN to reconstruct images [17]. Traditional methods such as these based on network structure innovation are devised mainly to restore degraded images with a single mode, however, these methods are often ineffective in the restoration of degraded images with complex patterns.
Some exploratory work has been carried out in the field of deep neural networks for complex tasks in low-level vision. Among these methods, representative examples are CNN denoising prior [18] and DualCNN [19]. Both the CNN denoising prior and DualCNN methods have achieved good results in low-level visual tasks, such as image denoising, deblurring and super-resolution. However, since the degraded data for network training are independent of each other, the trained network can only solve a single task at a time. In actual image restoration such as spatial target recognition, degradation is often caused by a variety of complex factors, especially those based on ground-based observations. Degraded images are not only affected by atmospheric turbulence during imaging but are also susceptible to interference from ground gas, stray light, high-energy particles, nebulae, thermal noise and electronic noise. Thus, images ground-based spatial targets are not only susceptible to severe degradation but also can have very low peak signal-to-noise ratios (PSNR) [20], [21], [22].
The traditional solution to such complex low-level vision tasks is to improve the model’s presentative ability by increasing the complexity of the model. However, more complex models often need to be provided with more parameters and are, therefore, more prone to neural network overfitting, a challenge that several researchers have worked hard to resolve. Many functions [23] have been used to characterize the generalization boundary. Many of these bounds are obtained either through some forms of regularization (typically L2 regularization [24], [25]) or by restricting the complexity of the function class. Mainstream strategies for preventing overfitting during neural network training include drop-out, early stoppage and data augmentation [23]. Recently, the network pre-training method [26], the network layered-training method [27] and the method of using auxiliary variables for hidden layers [28] have also been proposed.
The reconstruction of complex low-level visual tasks, especially the restoration of degraded images caused by multi-factor coupling, is very difficult. Compared to image degradation caused by a single factor (such as motion blur), the patterns in image degradation caused by multi-factor coupling are more difficult to learn, and the degraded spaces tend to be larger, thus making restoration more problematic. The end-to-end deep neural network is not conducive to learning the inverse transformation mode of image degradation when training data is limited, resulting in a poor quality of restoration. Therefore, this study proposes an optimization strategy based on the regularization and decomposition of complex tasks for the blind removal of turbulence blur, based on the characteristics of degraded images caused by multi-factor coupling, together with the concept of complex task decomposition. The results of this strategy show it to be more effective than other existing methods, and the main contributions of this paper are threefold, namely:
(1) A general regularization decomposition strategy for complex tasks is proposed, in which different parts of the deep neural network can focus on distinct aspects of the entire problem (the different parts of the network correspond to different parts of the problem), thus achieving task decomposition and reducing the complexity of the problem.
(2) Based on the characteristics of image degradation caused by multi-factor coupling in the actual imaging process, an optimization strategy based on regularization decomposition for complex tasks is designed for use in the restoration of atmospheric turbulence-blurred images. The results of the image restoration effect of the proposed strategy are superior to those of other existing methods.
(3) This method is a regularization strategy based on task decomposition, designed to reduce the complexity of the problem, enhance the generalization of the network, and reduce overfitting.
The remainder of this paper is organized as follows: Section 2 provides an overview of related research work; Section 3 analyzes the motivation and rationale behind the proposed method in a physical sense; Section 4 details the design and validity analysis of the TDROS; Section 5 combines the TDROS with a restoration model for turbulence-blurred images; Section 6 describes the experiments performed on the proposed model and compares the results with those of some existing models; and, finally, Section 7 summarizes and provides conclusions to the work.
Section snippets
Related works
Restoration of complex images. Several pioneering studies have investigated multiple simultaneous degradations. Gao et al. [29] used a convolutional autoencoder to blindly deconvolve a single-frame spatial target image, while Kim et al. [11] used a 20-layer deep CNN network to solve the super-resolution of multi-scale images. Guo et al. [30] built a one-to-many network that can handle images with different levels of compression artifacts. Zhang et al. [31] proposed a 20-layer deep CNN to
Principle analysis of the effectiveness of optimization with complex task decomposition and regularization
Spatial target image restoration and reconstruction based on ground observation is known to be problematic. In particular, it is difficult to achieve good results with a traditional single-layer optimization model. Deep learning models used for image restoration have weak general capabilities, and deep learning models trained with limited data sets will not be universally applicable to diverse natural images. Nonetheless, due to the unique characteristics of spatial target observation, we
TDROS application model
Combined with the concept put forward in Section 3.3 above, the TDROS application model to address the reconstruction of severely degraded images is proposed, as shown in Fig. 1, which can be applied to any scenario with decomposed complex tasks. After the TDROS is introduced into the deep convolutional neural network, the network is divided into two parts, the front of which gives the network the ability to perform major tasks as well as potential tasks, while the latter part of the network is
Proposed method for turbulence-degraded images
Based on the previous research results, this paper used the TDROS to design a novel deep neural network for the restoration of atmospheric turbulence-degraded images.
Comparative analysis of experiments
For the restoration of severely degraded images of spatial targets, no effective objective evaluation metric can be used [58]. Therefore, the reference image evaluation metric was used in the evaluation of the experimental results of the simulated degraded test images, and a subjective evaluation by human vision was adopted for the evaluation of the experimental results of the real degraded images of the spatial target. To better analyze the performance of the proposed algorithm, several
Conclusions
In this paper, we propose a general complex task decomposition regularization optimization strategy (TDROS) for image restoration. This strategy can be flexibly applied to existing neural networks to solve complex image restoration problems, thereby effectively improving the generalization ability of the network. The experimental analysis showed improved performance and generalization ability in neural networks enhanced by the introduction of the TDROS.
Inspired by the success of the TDROS, a
Declaration of Competing Interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
Acknowledgments
This work has been partially supported by the Key scientific research fund of Xihua University, China (Grant No: Z17134), Sichuan science and technology program, China (Grant No: 2020YFG0188, 2021YFG0022).
References (58)
- et al.
Stacked convolutional auto-encoders for single space target image blind deconvolution
Neurocomputing
(2018) Learning from hints in neural networks
J. Complexity
(1990)- et al.
Semi-supervised change detection method for multi-temporal hyperspectral images
Neurocomputing
(2015) - et al.
Progress in the restoration of image sequences degraded by atmospheric turbulence
Pattern Recognit. Lett.
(2014) - S. Nah, T. Hyun Kim, K. Mu Lee, Deep multi-scale convolutional neural network for dynamic scene deblurring, in:...
- et al.
Online monitoring of flotation froth bubble-size distributions via multiscale deblurring and multistage jumping feature-fused full convolutional networks
IEEE Trans. Instrum. Meas.
(2020) - et al.
Velocity estimation from a single linear motion blurred image using discrete cosine transform
IEEE Trans. Instrum. Meas.
(2019) - et al.
A deconvolutional reconstruction method based on Lucy–Richardson algorithm for joint scanning laser thermography
IEEE Trans. Instrum. Meas.
(2021) - J. Zhang, J. Pan, J. Ren, Y. Song, L. Bao, R.W. Lau, M.-H. Yang, Dynamic scene deblurring using spatially variant...
- S. Lefkimmiatis, Non-local color image denoising with convolutional neural networks, in: Proceedings of the IEEE...
A novel fast single image dehazing algorithm based on artificial multiexposure image fusion
IEEE Trans. Instrum. Meas.
Image super-resolution using deep convolutional networks
IEEE Trans. Pattern Anal. Mach. Intell.
Learning to deblur
IEEE Trans. Pattern Anal. Mach. Intell.
Deep convolutional neural network for image deconvolution
Image restoration using very deep convolutional encoder-decoder networks with symmetric skip connections
Removing atmospheric turbulence via space-invariant deconvolution
IEEE Trans. Pattern Anal. Mach. Intell.
Satellite image enhancement and restoration - a review
Int. J. Comput. Sci. Commun. Netw.
Weighted low-rank tensor recovery for hyperspectral image restoration
IEEE Trans. Cybern.
Supervised autoencoders: Improving generalization performance with unsupervised regularizers
Stability and generalization
J. Mach. Learn. Res.
Covering number bounds of certain regularized linear function classes
J. Mach. Learn. Res.
Cited by (2)
- ☆
This paper has been recommended for acceptance by Zicheng Liu.