图像作为视觉信息的主要载体，是人们高效准确获取信息的重要保证。然而在雾霾天气条件下，户外成像系统采集的图像存在严重的退化，导致出现对比度下降、边缘细节丢失以及图像变暗等问题。针对上述问题，在分析现有算法的基础上，本文从图像结构、图像颜色和图像亮度三个角度对现有去雾算法进行优化，并扩展应用到交通安防场景图像处理中。本论文主要内容和贡献如下：1）针对雾天图像在结构上丢失边缘信息和细节信息的问题，本文提出了一种基于结构先验补偿的图像去雾方法。该算法以PeleeNet为基础设计了一个生成式对抗网络结构，并将图像的高低频信息作为附加的先验输入网络中，使网络更加关注图像结构特征的学习。在公开的雾天图像数据集RESIDE上与相关去雾算法进行客观指标和视觉效果比较，该算法在PSNR和SSIM指标上相比FD-GAN模型的结果分别提升了18.1%和4.9%，同时能够保证有雾图像细节和纹理轮廓的恢复。2）针对雾天图像在色彩上出现对比度下降、饱和度弱化的问题，本文提出了一种基于彩色空间模型解耦的图像去雾算法。该算法将HSV颜色空间下有雾/无雾图像的差异以对比正则化的方式设计为颜色先验损失函数，以求模型达到在图像色彩上的解耦，同时添加了CBAM注意力机制以增强模型对通道和空间信息的关注。实验表明，该算法在保留结构信息的同时能够更好恢复色调、饱和度等颜色信息，在PSNR和SSIM指标上分别取得35.10dB，0.9897，优于其它对比模型。此外，颜色先验损失对提升模型的去雾和重建效果较好，其中PSNR提升了6.05%，SSIM提升了3.2%。3）针对现有去雾算法结果存在的图像整体亮度偏暗的问题，本文提出了基于亮度统计先验的雾天图像增强方法。该方法基于Retinex理论，利用良好天气图像V通道统计先验对光照分量进行伽马增强。同时对饱和度通道S进行补偿以避免出现颜色失真。实验表明，该算法能有效提高亮度信息，在PSNR和SSIM指标上相比RetinexNet算法分别提升了9.5%和16.8%。论文用计算机技术提高了雾霾天气下采集图像的视觉可见性，并将多种算法集成到交通场景图像去雾与增强系统设计中，体现了学科的交叉与应用性。
As the main carrier of visual information, images are an important guarantee for people to obtain information efficiently and accurately. However, under hazy weather conditions, the images captured by outdoor imaging systems suffer from serious degradation, leading to problems such as contrast degradation, edge detail loss and image darkening. To address above problems, based on the analysis of existing algorithms, this paper optimizes the image dehazing algorithm from three perspectives: image structure, image color and image brightness, and extends the application to image processing of traffic security scenes. The main contributions are as follows:1) To address the problem of losing edge information and detail information in the structure of hazy images, this paper proposes an image dehazing method based on structure prior-based compensation. The algorithm designs a generative adversarial network structure based on PeleeNet and inputs the high and low frequency information of images as additional priors into the network to make the network more concerned with learning the structural features of images. We performed objective metrics and visual comparisons on the publicly available RESIDE dataset, and the method achieved 18.1% and 4.9% improvement in PSNR and SSIM metrics over the FD-GAN model, while being able to ensure the recovery of details and texture contours of hazy images.2) To address the problem of visual contrast degradation and saturation weakening in hazy images, this paper proposes an image dehazing method based on the decoupling of color space model. The algorithm designs the difference between hazy/ haze-free images in HSV color space as a color prior loss function with contrast regularization in order to achieve color decoupling. And then we add a CBAM attention mechanism to enhance the attention of the model to channel and spatial information. Experiments show that the algorithm can better recover color information such as hue and saturation while preserving structural information, and achieves 35.10 dB and 0.9897 in PSNR and SSIM metrics, respectively, which is better than other contrast models. In addition, the visual prior loss is more effective in enhancing the dehazing and reconstruction of the model, in which PSNR is improved by 6.05% and SSIM is improved by 3.2%.3) To address the problem of overall image brightness darkness that exists in the results of existing dehazing algorithms, this paper proposes a hazy sky image enhancement method based on brightness prior. The method is based on Retinex theory, and adaptive gamma enhancement is performed on the light component using the statistical prior of V channel of images in good weather. At the same time, the saturation channel(S) is compensated to avoid color distortion. Experiments show that the algorithm can effectively improve the luminance information, with 9.5% and 16.8% improvement in PSNR and SSIM metrics, respectively, compared to the RetinexNet algorithm. In this paper, multiple algorithms are integrated into the traffic scene image dehazing and enhancement system to realize the practical scene application of the algorithm. Our work uses computer technology to improve the visual visibility of images captured in hazy weather, and multiple algorithms are integrated into the traffic scene image dehazing and enhancement system, and reflects the intersectionality and applicability of the discipline.
