【作者】 张敏；

【导师】 王园宇；

【作者基本信息】 太原理工大学 ， 计算机科学与技术， 2017， 硕士

【摘要】 人类的生产生活对自然环境造成了严重影响,导致雾霾天气出现的日渐频繁。雾天气条件下拍摄的照片模糊不清,对各类视频监控系统的使用带来了不便,因此对图像进行去雾处理,恢复出清晰的图像,是现在的研究热点,也是很有实际价值的课题。根据颜色衰减先验知识进行去雾是近两年提出的一个新的去雾方法,该方法快速有效,极大的提高了去雾速度,具有推广价值。笔者在实验过程中发现,该算法中大气散射系数的选择对最终的去雾结果影响很大,需要进行人工调整。而且,颜色衰减先验知识,在雾浓的区域失效,因此对雾浓的图像处理效果不佳。本文针对这两个问题进行了改进。本文在颜色衰减先验知识的基础上,提出了基于小波融合的图像去雾算法。首先,建立了透射率关于图像亮度、饱和度的线性模型。其次,选用400幅图像及其准确的景深信息用于训练样本的生成,并将大气散射系数的概率分布直接融入到训练样本中,保证了训练样本生成的准确性,随后采用机器学习中的监督学习算法估计出图像透射率,然后用小波融合算法将估计出的透射率与图像灰度图的反转图像融合。最后,用细化后的透射率信息对图像进行去雾处理。直接对透射率进行建模的过程,避免了大气散射系数的选择,用图像灰度图的反转图对透射率进行细化,提高了透射率的准确性,最终提高了对雾浓图像的处理效果。实验结果证明本文算法可行有效。此外文章采用通用的去雾图像质量评价方法将本算法与先进的去雾算法进行了比较。实验结果表明,本文算法提高了透射率的准确性,改善了浓雾区域的恢复效果。具有适用性好,计算复杂度低的特点。本文的主要贡献为以下几点:(1)建立了透射率关于图像亮度饱和度的线性模型,避免了去雾过程中对大气散射系数的人工选择。(2)得到大气散射系数的分布直方图,并将其代入训练样本中,从而获得准确度高的训练样本,保证训练出模型的可靠性。(3)提出图像灰度图的反转图能够作为透射率对图像进行去雾处理,尤其对雾浓的图像,处理效果较佳。(4)利用小波算法将估计出的透射率与图像灰度图的反转图融合,提高透射率的准确性,保证最终的去雾效果。更多还原

【Abstract】 Human production and life have badly effect on the natural environment,which leading to fog haze weather appears frequently.The pictures taken under foggy weather are blurred.Haze images affect the use of various video surveillance systems.Therefore,to recover clear images is now a hot research topic and also a topic full of practical value.Single image haze removal using color attenuation prior is a new method proposed in recent years.The method is fast and effective,and greatly improves the speed of fog removal,and has promotional value.However in the algorithm,the choice of atmospheric scattering coefficients has a great influence on the final fog removal results and manual adjustment is required.And color attenuation prior works badly in thick fog area so that this algorithm can not restore a clear image when there is thick fog in the image.In this thesis,these two problems have been improved.A single image dehazing algorithm by wavelet fusion based on color attenuation prior is proposed in this thesis.First of all,a linear model of relationship between transmission and brightness and saturation is built.Secondly,400 images and their accurate depth of field information are used to generate the training samples.The probability distribution of atmospheric scattering coefficients is directly incorporated into the training samples,which guarantees the accuracy of training samples.And the supervised learning algorithm in machine learning is used to estimate the image transmission.Then the wavelet fusion algorithm is used to fuse the estimated transmittance with theinverse image of the grayscale image.Finally,the fog is processed by the optimized transmittance information.Making a model of transmittance can avoid the choice of atmospheric scattering coefficient.The transmittance refined by the inverse image of the grayscale image can improve the accuracy of transmittance.Improve the performance of this algorithm in thick fog area.Experimental results show that the algorithm proposed in this thesis is feasible and effective.At the end of the thesis,several common quality evaluation methods of dehaze are used to compare the performance of the algorithm in the thesis and the classical algorithms.The experimental results show that the algorithm can achieve better results for all kinds of fog images with a short time.The main innovations of this thesis are summarized as follows:(1)A linear model of transmittance about image’s brightness and saturation is built which avoids the manually selecting of the atmospheric scattering factor.(2)The histogram of the atmospheric scattering coefficient is obtained and put into the training samples,which helps to obtain training samples with high accuracy and ensures the reliability of the model.(3)The inverse image of the gray scale image can be used as the transmittance to haze removal.And it works well in thick fog area.(4)The wavelet transform is used to fuse the inverse information of the gray scale image and the rough depth map to improve the accuracy of the transmission and ensure the quality of haze removal.更多还原