具有透射率先验与亮度感知的图像去雾算法

Image defogging algorithm with transmittance prior and luminance awareness

在图像去雾领域,图像明亮区域的去雾效果与去雾后图像的抗噪性能是重要的去雾评估标准,然而大多数去雾算法在处理有雾图像明亮区域时存在一定程度的失真问题,并且得到的去雾后图像常存在一定噪声,导致去雾后图像的抗噪性能普遍较弱。针对以上去雾问题,文中提出了具有透射率先验与亮度感知的图像去雾算法(Image Dehazing Algorithm with Transmittance Prior and Luminance Awareness,DTPLA)。该算法首先构造求解大气光值模块(Solving Atmospheric Light Value Module,SALVM),将该类问题转化为克服干扰因素的影响和缩小求解误差的问题;其次,提出透射率先验(Transmittance Priori,TP)方法,以对图像明亮区域透射率的取值问题进行优化求解,并设计图像亮度感知模型(Image Luminance Perception Model,ILPM),对图像可视化效果进行增强,通过对透射率取值与可视化效果的联合优化解决失真问题并丰富图像重要信息;最后,引入图像降噪模块(Image Noise Reduction Module,INRM),将其嵌入到图像去雾模型中,以提升去雾后图像的抗噪性能;仿真结果表明,与其他主流算法相比,运用所提算法得到的去雾结果在3种数据集下均具有更好的性能表现,图像的MSE值、PSNR值、SSIM值分别平均达到7.4701,39.9839 dB,83.72%,基于以上3种数据集的实验结果验证了所提算法的有效性。

Objective Haze is extremely common in our daily lives and is a common natural phenomenon. In recent years,the presence of atmospheric haze has had a serious impact on our production and life, making our daily travel extremely inconvenient and limiting people's outdoor activities. For vision-based intelligent machines, the presence of haze seriously reduces the quality of images captured by intelligent devices. In recent years, a variety of haze removal methods have appeared, such as image enhancement, image restoration, deep learning based haze removal methods, but the use of the above haze removal methods in the image of the bright areas of the processing effect is not ideal, prone to distortion. Meanwhile, the use of the above haze removal methods to remove the haze of the image obtained after the PSNR performance is generally lower, indicating that after the removal of the haze of the image of the presence of the corresponding noise, and has not been reasonably removed. The mainstream defogging methods can not solve the above two problems at the same time. Therefore,it is necessary to deeply process the haze present in the image.Methods Considering the shortcomings of the current defogging methods, based on the atmospheric scattering model and combined with the dark channel prior theory, we propose an image defogging algorithm with transmittance prior and luminance perception(Fig.1). The algorithm first constructs a module for solving the atmospheric light value, which transforms this type of problem into a problem of overcoming the influence of interference factors and reducing the solution error;Secondly, it introduces a Gaussian filtering noise reduction module, which is embedded into the image de-fogging model in order to improve the anti-noise performance of the de-fogged image;Finally, it proposes a transmittance a priori method in order to correct the transmittance of the bright region, and designs the image luminance sensing model(Fig.2-3) to enhance the visualization effect of the image, and the image visualization effect is enhanced.Results and Discussions The experiments are carried out under three datasets: the synthetic foggy dataset, the SOTS dataset, and the HSTS outdoor dataset, in order to verify the effectiveness of the proposed method.According to the performance evaluation results of this method and the mainstream algorithms on the synthetic dataset(Fig.7), the degree of distortion and the degree of color shift of the de-fogged image obtained by this algorithm is lower than that obtained by other algorithms, and more important information in the image can be retained. The average PSNR value of the image obtained by the algorithm in this paper is 39.579 9 dB, which has a significant advantage over other defogging algorithms, indicating that the image has a better anti-noise performance and lower noise content. From the results of image SSIM evaluation, the defogging performance obtained by using the algorithm in this paper reaches 82.54%, indicating that the defogging results obtained by using the algorithm in this paper are comparatively more similar to the original image, and can maximize the retention of important information in the image. According to the performance evaluation results of the performance evaluation on the SOTS dataset(outdoor), it can be seen(Fig.8) that the average value of the image PSNR obtained by applying the algorithm of this paper is 39.601 8 dB and the average value of the SSIM reaches82.84%. According to the performance evaluation results of the performance evaluation on the HSTS dataset(Fig.9), the average value of image PSNR obtained by using this paper's algorithm is 40.770 1 dB, and the average value of image SSIM reaches 88.78%. The experimental results under the three foggy datasets all show that the algorithm in this paper has a significant advantage over other mainstream de-fogging algorithms, which not only can effectively deal with bright regions such as the sky, but also can remove the noise content, which verifies the effectiveness of the proposed algorithm.Conclusions As can be seen from the experimental results(Tab.7), the use of Ref.[22] algorithm to get a corresponding good de-fogging effect, but also consumes a relatively large amount of de-fogging time, the time cost is larger, in the requirements of the de-fogging efficiency of the scene can not be well applied. The use of Ref.[23] algorithm to get the de-fogging effect is more ideal, but comprehensively, in the de-fogging performance and de-fogging efficiency is slightly weaker than the algorithm proposed in this paper;The use of Ref.[12]algorithm for the de-fogging processing of the time cost is small, but the de-fogging effect is not ideal.Comprehensively, the use of the algorithm proposed in this paper in the fogging effect and time cost have good performance.