针对油气长输管道采用无人机巡检时所拍摄的红外图像去模糊问题，利用图像通道的先验知识提升模糊图像质量，分别基于双边滤波和非盲去模糊网络NBDN去除人工伪影的方法达到更佳的图像复原效果。首先，基于暗通道先验知识，在最大后验的优化框架中添加暗通道的L₀正则项；然后使用图像梯度的L₀正则项，代替图像像素的L₀正则项作为潜在图像的正则化约束，使用迭代交替估计图像模糊核和中间潜在图像；采用半二次分裂方法和查表法间接优化求解，估计中间潜在图像；采用双线性插值估计图像模糊核，通过对图像进行上下采样，构建图像金字塔，进而利用共轭梯度法直接优化求解。最后，利用估计的模糊核，使用基于超拉普拉斯先验的图像非盲去模糊方法得到潜在图像I₁；使用基于L₀正则化的非盲去模糊方法得到潜在图像I₀；计算估计的潜在图像I₁和I₀之间的差值映射，从I₁中减去双边滤波过滤后的差分图，得到最终的潜在图像I。将本文算法在低照度图像、含有饱和像素的图像、真实图像以及红外摄像图等图像数据上进行实验，相对于其他图像去模糊算法。实验结果表明：所提出的方法在多种模糊图像复原效果上均具有较强的竞争力。

A novel method was proposed to enhance the quality of blurred infrared images captured during unmanned aerial vehicle (UAV) inspections of oil and gas pipelines. The issue of image deblurring was addressed by utilizing prior knowledge of image channels and employing bilateral filtering and the non-blind deconvolution network (NBDN) to remove artificial artifacts. Firstly, the dark channel prior knowledge was incorporated into a maximum a posteriori optimization framework by adding a dark channel L₀ regularization term. Then, instead of using L₀ regularization on image pixels, the L₀ regularization term based on image gradients was employed as the constraint for the latent image. The blur kernel and the intermediate latent image were iteratively estimated through alternating estimation techniques and indirect optimization methods including semi-quadratic splitting and table lookup. The blur kernel was estimated using bilinear interpolation, and an image pyramid was constructed by upsampling and downsampling the image, which were then directly optimized by the conjugate gradient method. Finally, with the estimated blur kernel, a non-blind deblurring method based on the super-Laplacian prior was presented to obtain the latent imageI₁, while another non-blind deblurring method based on L₀ regularization was also applied to obtain the latent image I₀. The difference map between I₁ and I₀ was calculated and then subtracted from I₁ by bilateral filtering to obtain the final latent image I. The experiments were designed on low-light images, images with saturated pixels, real images, and infrared camera images to asses the proposed algorithm. The results show that the proposed method has strong competitiveness in various blurry image restoration effects.