为了提高影像阴影去除的效果，提出一种基于最小噪声分离（MNF）和生成对抗网络（GAN）的影像阴影去除算法。它以GAN作为基本框架，在生成器和判别器中分别引入条件信息，采用端到端共同学习的多任务模式。生成网络采用编码解码结构，判别网络采用马尔可夫判别器结构。此外，此算法使用MNF，将消除噪声的影像灰度化后与阴影影像一起训练，进而恢复无阴影的影像。这样的网络在训练时可以专注于MNF变换后的单独特征嵌入，而非传统的跨任务共享嵌入。实验结果表明，在指定数据集上，所提算法的结构相似性（SSIM）的平均值达0.9780，像素均方根误差（RMSE）的平均值减小到9.8717。在主观感知和客观评价指标上，所提算法的实验结果均优于对比算法。

An image shadow removal algorithm based on minimum noise fraction (MNF) and a generative adversarial network (GAN) is proposed to improve the shadow removal effect. The algorithm takes GAN as its basic framework, introduces condition information into the generator and discriminator respectively, and adopts the multitask mode of end-to-end joint learning. The generative network adopts the encoding-decoding structure, and the discriminant network adopts the Markov discriminator structure. Additionally, the proposed algorithm uses MNF to restore the shade-free image after graying the noise-eliminating image with the shadowed image. Therefore, our network can focus on single feature embedding after the change in MNF instead of the traditional cross-task shared embedding. Experimental results indicate that the proposed algorithm can increase the mean structural similarity (SSIM) to 0.9780 and decrease the mean root mean square error (RMSE) to 9.8717 on the specified dataset. Both visual and statistic comparisons confirm that the proposed algorithm is better than other algorithms.