Effective haze removal under mixed domain and retract neighborhood
Introduction
The quality of images would exhibit degraded visibility during inclement weather conditions. Therefore, outdoor scenes usually have low visibility and contrast, which lead to depressed reliability of many vision systems such as the automatic vehicle driving system and the intelligent navigation system. Haze removal just decreases the influence of the haze and tries to restore such this image by recovering the aesthetic quality and also improving the details (Fig. 1).
Normally, numerous dehazing approaches are based on physical degradation model. It can be considered as a linear combination of the clear scene and the air-light, as shown in Eq. (1): where c ∈ {r, g, b}, Fc(x) ∈ RC is the haze image, Jc(x) is the clear scene radiance, Ac is the brightest point in the sky, and t(x) is the scene transmission with the depth. Since Ac is usually obtained from experience, t(x) has already become the most important research in haze removal. To make the ill-posed problem tractable, the auxiliary depth information is an effective way for solving this problem. However, not all scenes have objective depth information. Different from early algorithms, most existing endeavors are devoted to finding strong assumptions or constraints for overcoming disadvantage influence. So far as known, there are two kinds of statistical methods for estimating the transmission map. The first type is called Contrast prior [1], [2], [3], [4], [5] which has been taken by Tan [2] as the representative method. Contrast [4] and edge information [5] are treated as an effective clue so that optimal transmission map is dynamically calculated. Another class of techniques is Chromaticity prior [6], [7], [8], [9], [10], [11], [12], [13], [14], [15], [16], [17], [18], [19], [20]. It has been demonstrated to be widely attractive and practical than contrast prior. Assuming that similar color has the same scene depth, color consistency can be regarded as an indication for media transmission map [6]. For instance, Fattal [7] exploits the geometrical model to restore degraded images. This is accomplished via the assumption that the attenuation factor and surface shading are locally independent. It can produce positive dehazing images but complex computation has been introduced. Based on the theory of Color Line, Fattal [8] also uses geometric triangle to express the relationship of local pixels and obtains the scene transmission map. As the most concerned method, He et al. [9] firstly discovers a simplest dark channel prior that a minimum intensity value in a haze free image will be found in at least one color channel. In consideration of this prior, the initial transmission can be easily predicted. Due to the intelligible theory and satisfactory performance, many progressive methods [10], [11], [12], [13], [14], [15], [16], [17], [18], [19], [20], [21], [22], [23] have been proposed for variety applications on the foundation of DCP. For example, in order to refine the preliminary results, Tarel and Hautiere [10] infers a veil refined by median filter, Jing et al. [11] and He et al. [12] employ bilateral and guided filter to smooth the atmospheric veil's rough estimation, respectively. In view of the boundary constraint theory, Lai et al. [13] and Wang et al. [14] restored better results than traditional dehazing methods. Unfortunately, halo artifacts and color distortion are still introduced in the sky region which does not conform to DCP constraint. As a result, Wang et al. [15] and Luan et al. [16] put forward the methods to predict the dynamic atmospheric light separately. Besides that, ground radiance suppressed-HTM [17], HDCP module [18] and VR module [19] are employed to directly improve DCP prior. Laplacian-based restoration [20] and linear intensity transformation [21] are also adopted to restore the degradation images.
Note that although all of DCP-based methods have obtained better restoration results, these algorithms still have common problems. As a matter of experience, one of the main reasons is that the principle of DCP. It makes the restored image darker and details that cannot be distinguished. More importantly, it is obvious that the process of light scattering not only changes image luminance but also cuts saturation down. Nevertheless, most of algorithms are only based on the luminance characteristics and neglect the effect of the saturation layer of images.
To resolve the aforementioned problems and motivated by HSI color space, we improve the DCP estimation by mixed strategy for dehazing single images. Unlike traditional models for estimating the transmission map according to metabolic brightness, we combine the saturation with luminance layers, and perform to estimate the initial transmission map. Similar to other methods on suppressing the influence of halo artifacts and inaccuracy points in the initial transmission map, a refinement of process is introduced to achieve the adaptive transmission compensation. Considering that the local color priors have an influence on the accuracy of the recovery results, we utilize local color constraint to generate the optimal transmission map based on the calculation of geometric projection. The main progresses of our proposed approach are organized into two parts as follows:
- (1)
Different from many others dehazing techniques which only focus on the luminance layer assumption for estimating the initial transmission map, our method emphasize more on the combination of the saturation and luminance layer. As a result, the proposed can produce high quality fog free image effectively which more visually attractive to observers.
- (2)
Compared with the color line model often discards patches [8], we assign essential conditions to shrink the patch for finding out the valid points around the center pixel. More importantly, we deploy them as a constraint for edge preserving, which brings down the error rates of Color Line.
Experiments have shown the proposed method performs on par with or better than the advanced dehazing methods. Firstly, we brief dehazing approaches can be treated as a typical situation of image restoration so that our visual features can restore the initial transmission map through the variational framework in Section 2. In Section 3, we presents improved color line model which is dedicated for producing refined and regularized transmission maps for dehazing algorithm in detail. Furthermore, we present plenty of comparison experiments to verify the effectiveness of our algorithm in Section 4. And finally, we conclude the paper and the future work is given in Section 5.
Section snippets
Model analysis
In this section, through theoretical and graphical analysis in the RGB and HSI model, we can clearly find out the solution of the transmission map. Moreover, we also serve the derived representation as energy cost function for achieving an accurate solution.
At present, a physical degradation model is usually modeled by Eq. (1). Most of approaches have been demonstrated to be successful based on atmospheric scattering model. However, many methods are only one-sided pursuit the improvement of the
Refined estimation
Similar with HE algorithms [9], [12], the initial transmission should be smoothed into refined t(x) by minimizing the cost function. Obviously, HE [9] regularized the soft matting and enforced smoothing weighted by λ.
To further speed up, Guided Filter [12] can be used to improve the rough transmission map: where |ω| is the pixel number in the patch ωk, and . Here, ak
Comparison experiments
In order to obtain the more objective results, numerous experiments were conducted. We employ synthetic and real test database [27], [28] to evaluate our method. The images are collected by: http://perso.lcpc.fr/tarel.jeanphilippe/, http://www.cs.huji.ac.il/∼raananf/projects/dehaze_cl/results/.
Conclusion
Based on mixed strategy, we present a method which can obtain an optimal transmission map and restore the haze free image. In order to obtain the better initial transmission, we combine the luminance and saturation layer together. Afterwards, we use the local color and gradient as constraint, which can derivate an accurate transmission map. The experimental results show the method is effective to achieve well representation. The main problem of this algorithm is that the parameters cannot
Acknowledgments
This work was supported by the National Natural Science Foundation of China, under Grants 61701524, 61372167 and 61379104. Thanks to the anonymous reviewers for their valuable suggestions.
Linyuan He received the B.S. and M.S. degrees in electrical engineering from Air Force Engineering University, Xi'an China, in 2005 and 2008, respectively. Both in electrical engineering, he is currently pursuing the Ph.D. degree in electrical engineering in Xi'an Jiao tong University. He is also a lecturer in the School of aeronautical engineering, Air Force Engineering University, shanxi, China. His research interests include computer vision, image processing and Deep Learning.
References (36)
- et al.
Perception oriented transmission estimation for high quality image dehazing
Neurocomputing
(2017) - et al.
Dehazing for images with large sky region
Neurocomputing
(2017) - et al.
Haze removal for a single visible remote sensing image
Signal Process.
(2017) - et al.
Single satellite image dehazing via linear intensity transformation and local property analysis
Neurocomputing
(2016) - et al.
Contrast restoration of weather degraded images
IEEE Trans. Pattern Anal. Mach. Intell.
(2003) Visibility in bad weather from a single image
- et al.
Effective contrast-based dehazing for robust image matching
IEEE Geosci. Remote Sens. Lett.
(2014) - et al.
Edge collapse-based dehazing algorithm for visibility restoration in real scenes
Jo. Disp. Technol.
(2016) - et al.
Single image defogging by multi-scale depth fusion
IEEE Trans. Image Process.
(2014) Single image dehazing
Dehazing using color line
Single image haze removal using dark channel prior
Fast visibility restoration from a single color or gray level image
Physics-based fast single image fog removal
Guided image filtering
IEEE Trans. Pattern Anal. Mach. Intell.
Single-image dehazing via optimal transmission map under scene priorss
IEEE Trans. Circuits Syst. Video Technol.
Fast image dehazing method based on linear transformation
IEEE Trans. Multimed.
Haze removal method for natural restoration of images with sky
Neurocomputing
Cited by (11)
NLDN: Non-local dehazing network for dense haze removal
2020, NeurocomputingCitation Excerpt :In [22], Tang et al. seek a best combination of local maximum contrast, dark channel prior, hue disparity and local maximum saturation using forest regression to estimate transmission map. A multi-scale fusion scheme [23] for single image dehazing is proposed. Zhu et al. [11] propose a linear model to create the relation between the scene depth of the hazy image under color attenuation prior and solve the model with a supervised manner.
Single image haze removal using contrast limited adaptive histogram equalization based multiscale fusion technique
2024, Multimedia Tools and ApplicationsVLSI Design of Saturation-Based Image Dehazing Algorithm
2023, IEEE Transactions on Very Large Scale Integration (VLSI) SystemsHaze-line optimization image dehazing algorithm based on non-local prior
2023, Guangdianzi Jiguang/Journal of Optoelectronics LaserSingle Image Dehazing Using Saturation Line Prior
2023, IEEE Transactions on Image ProcessingImplementation of Satellite Road Image Denoising using Iterative Domain Guided Image Filtering with Gray World Optimization
2022, Journal of Communications
Linyuan He received the B.S. and M.S. degrees in electrical engineering from Air Force Engineering University, Xi'an China, in 2005 and 2008, respectively. Both in electrical engineering, he is currently pursuing the Ph.D. degree in electrical engineering in Xi'an Jiao tong University. He is also a lecturer in the School of aeronautical engineering, Air Force Engineering University, shanxi, China. His research interests include computer vision, image processing and Deep Learning.
Jizhong Zhao received the B.S. and M.S. degrees in mathematics and Ph.D. degree in computer science with a focus on distributed systems from Xi'an Jiao tong University, Xi'an, China, in1992, 1995, and 2001, respectively. He is a Professor with the Computer Science and Technology Department, Xi'an Jiao tong University. His research interests include computer software, pervasive computing, distributed systems, network security. Dr. Zhao is a member of the IEEE Computer Society and the Association for Computing Machinery (ACM).
Duyan Bi graduated from the Department of Electrical Engineering, National University of Defense Technology, chang'sha, China, in 1983, and received the M.S. degree in signal processing from National University of Defense Technology in 1987 and the Ph.D. degree in electrical engineering from Tours University, Tours, France, in 1997. He joined the School of aeronautical engineering, Air Force Engineering University in 1987, and is currently a Professor and the Director of the Laboratory of field reconnaissance and surveillance technology, Air Force Engineering University. His research interests include computer vision, pattern recognition and image processing.