均匀光照下单图像去雾新方法：基于复变滤波与精细传输估计

181 浏览量更新于2024-08-28 收藏 1.31MB PDF 举报

本文主要探讨了"光照不均匀影响下的单图像去雾方法"，由来自山东财经大学计算机科学技术学院的Wenhua Zhang、Shanshan Gao、Jing Chi和Hui Liu，以及山东大学计算机科学技术学院的Yuanfeng Zhou共同提出。在当前的研究背景下，由于在非均匀光照条件下估计大气值可能出现严重偏差，该论文创新性地引入了复卷积滤波技术来处理这个问题。论文的核心思想是首先利用复卷积滤波技术消除图像中的光照不均匀性，这有助于减少由于光照差异引起的伪像。通过这种方法，作者先对原始图像进行预处理，提取出更稳定的暗通道信息，为后续去雾过程提供基础。接着，他们应用了Super Pixel Segmentation（SLIC）算法对图像进行区域划分，以粗略估计传输率，这种方法提高了处理效率并减少了噪声。在优化图像传输过程中，论文提出了一种更为精确的传输率估计新算法，以抵消非均匀光照带来的影响。该算法旨在提高去雾效果的准确性，确保在复杂的光照环境下也能得到清晰的去雾结果。最后，作者可能还考虑了引导滤波技术的应用，它能够提供平滑而保持边缘细节的图像增强，进一步增强了去雾后的图像质量。这篇研究论文关注了单图像去雾任务中的一个关键挑战，即如何有效地应对光照不均匀性，通过结合复卷积、SLIC分割和精细的传输率估计，提出了一个系统性的解决方案，为实际的图像去雾应用提供了新的思路和技术支持。其研究成果对于改善低光照条件下的图像质量和视觉体验具有重要意义。

Single Image Dehazing Method Under the Influence

of Un-uniform Illumination*

Wenhua Zhang, Shanshan Gao, Jing Chi, Hui Liu

Department of Computer Science and Technology

Shandong University of Finance and Economics

Jinan, China

Shandong Provincial Key Laboratory of Digital Media Tech

nology, Jinan china

sxuzwh@163.com, gsszxy@aliyun.com

Yuanfeng Zhou

Department of Computer Science and Technology

Shandong University

Jinan, China

Abstract—Since the adverse effects may occur when the

atmosphere value under un-uniform illumination is estimated,

this paper proposed a dehazing method by using the

homomorphic filtering technology. The innovation of this

approach is that we use homomorphic filtering technology to

remove un-uniform illumination firstly before obtaining its dark

channel, then dehaze the pretreated image. During the process of

transmission optimization, applying SLIC super pixel

segmentation method to estimate the rough transmission rate,

then we use guided filtering for a more accurate transmission.

Besides, a new algorithm is proposed to estimate the atmosphere

value aiming to the affects causing by un-uniform illumination.

Experimental results show that this algorithm is efficient to avoid

the degradation of hazed images, thus the image clarity could be

improved.

Keywords—un-uniform illumination; dehaze; homomorphic

filtering; dark channel prior

NTRODUCTION

When taking pictures under the foggy weather, the image

tends to get affected and has low visibility. Distant objects lost

contrast in the fog, and its surroundings were also blurred. This

is mainly due to the scattering of atmospheric particles, which

lead to the object light attenuation. Thus, these objects will fade

and become similar to haze.

Early dehazing methods are mainly depended on the depth

of additional information or multi-observation of the same

scene. Classical methods are including [9], [3], [6], [7], [8] etc.

According to the different polarization characteristics of

scattered light, [9] used two or more same scene images with

different degree of polarization (DOP) to restore the scene

depth image to recover. Kopf et al [3] recommended use the

scene depth information of an image to dehaze. Narasimhan et

al proposed a model based on physical scattering [6]-[8]. With

this model, the structure of a scene can be easily recovered.

Single image dehazing, as its relatively few available

information of scene structure, becomes a more challenging

problem. Methods based on the atmospheric scattering physical

model are more popular, which treated haze and other

impurities as the main reason to image degradation, used the

concept of imaging under atmospheric environment to generate

degradation model according to atmospheric scattering. At

present, algorithms [1],[10],[2],[4] are more classic. Fattal[1]

estimate the scene transmission by considering the irrelevance

of surface projection and light transmission. However, this

method is based on the color statistics, so the haze image

processing effects is not capable. Tan [10] through maximum

the local contrast of haze images to achieve the purpose of

dehaze. The method can recover the details of the image well

in the process of dealing with the dense haze scene, but the

recovered image often appears the color saturation, and will

produce severe vignetting effect in the depth of field mutation

boundary. This is because the concentration of haze was

overestimated in the process to maximize contrast, so it does

not conform to the real physical model. Meng [2] established

the model by restricting the transfer function of the boundary

with context regularization weighted L1 norm, so as to estimate

the unknown scene transport optimization problem to achieve

the purpose of the demister. But when an image has large

regions of white, this method is difficult to identify whether

these pixels are haze coverage areas, and often leads to over

enhance the objects in the scene.

He [4] proposed a simple and effective method of dark

channel prior to dehaze. According to the prior: except the sky

area, most of natural images without haze blocks often contain

some brightness values within a small pixel in the color

channel, which is called the dark channel prior. Combining the

prior knowledge with soft matting methods, it can be simple

and efficient to recover the majority of haze pictures for high

quality. But when the object scene is very close to the

atmosphere optical properties, the algorithm is invalid.

Aiming at the shortcomings of He’s method, researchers

have proposed many improved method [12]-[17] based on the

algorithm. Pei et al [12] use a color transfer algorithm to

process the nighttime degraded images, it is also able to

improve the robustness of the dark channel prior by combining

with bilateral filtering algorithm. Feng et al [13] on the basis of

He’s algorithms, and combined with near-infrared image of the

scene, they utilize the dark channel prior to get the rough map

of each pixel, then optimize the transmission and dehaze by

applying Bayesian method.

*This work was supported by NSFC Joint Fund with Guangdong under Key

Project (No.U1201258), Natural Science foundation of China (No.61402261,

o.61303088, No.61572286, No.61332015), the scientific research

foundation of Shandong province of Outstanding Young Scientist Award (No.

BS2013DX048).