基于先验知识的图像快速去雾方法

A Method for Fast Image Dehazing Based on Prior Knowledge

摘要: 针对雾霾图像导致的视觉信息退化问题，本文提出一种基于先验知识的快速去雾算法。该方法通过图像分割技术精确估计大气光值，结合快速引导滤波与加权最小二乘联合滤波优化透射率图，有效抑制光晕伪影并保留边缘细节，并引入伽马校正与自动电平优化增强图像亮度对比度。实验采用包含天空区域/景深突变的室外图像、Middlebury_GT真实室内图像及Middlebury_Hazy雾化图像(通过Middlebury_GT深度图合成)构建多场景测试集。实验结果表明：在户外场景中，新可见边缘率(Er)为−0.0286，信息熵(IE)提升至7.6945，较颜色衰减先验的7.5850提升约1.44%，较原暗通道先验的7.5852提升约1.44%，较直方图均衡化的7.1655提升约7.38%；SSIM值为0.85，优于直方图均衡化的0.83。在室内场景中，新可见边缘率(Er)显著提升至3.6515，较原暗通道先验的2.1951提升约66.3%；信息熵(IE)为7.4335，SSIM值达到0.7786，较原暗通道先验的0.7587提升约2.62%，同时优于颜色衰减先验的0.6780和直方图均衡化的0.6719。

Abstract: To address the degradation of visual information in haze-affected images, this paper proposes a fast dehazing algorithm based on prior knowledge. The method employs image segmentation technology to accurately estimate atmospheric light values and combines fast guided filtering with weighted least squares joint optimization to refine the transmission map, effectively suppressing halo artifacts while preserving edge details. Gamma correction and automatic level optimization are introduced to enhance image brightness and contrast. Experiments utilize a multi-scenario test set comprising outdoor images with sky regions/depth mutations, Middlebury_GT real indoor images, and Middlebury_Hazy synthetic hazy images (generated from Middlebury_GT depth maps). Experimental results demonstrate that in outdoor scenes, the proposed method achieves a visible edge rate (Er) of −0.0286 and an information entropy (IE) of 7.6945, representing improvements of 1.44% over the color attenuation prior (7.5850) and the original dark channel prior (7.5852), and 7.38% over histogram equalization (7.1655). The SSIM value reaches 0.85, outperforming histogram equalization (0.83). In indoor scenes, the visible edge rate (Er) significantly improves to 3.6515, marking a 66.3% enhancement over the original dark channel prior (2.1951). The information entropy (IE) and SSIM values attain 7.4335 and 0.7786, respectively, surpassing the original dark channel prior (0.7587) by 2.62%, as well as outperforming the color attenuation prior (0.6780) and histogram equalization (0.6719). This work provides an efficient solution for computer vision tasks in complex hazy environments, effectively balancing edge preservation, detail recovery, and human visual perception characteristics.