基于条件生成对抗网络的图像去雾算法

Image Dehazing Algorithm Based on Conditional Generation against Network

本文改进了一种基于条件生成对抗网络的图像去雾算法。在生成器网络模型中使用32层Tiramisu代替U-Net可以减少训练参数,提高参数利用率。判别器网络的最后一层使用Sigmoid函数完成归一化。使用损失函数加权和的方式训练网络模型,使用一种评价去雾能力的分数,保存训练过程中分数较高的模型及参数,最后使用分数最高的网络模型及参数对室外真实图像进行去雾测试。训练集使用了室内和室外的真实无雾图像以及合成有雾图像,测试结果显示本文提出的模型和传统去雾方法相比,主观视觉效果、图像细节、色彩方面都有提高改善。

This paper improves an image dehazing algorithm based on conditional generation against networks. Using 32 Layer Tiramisu instead of U-Net in the generator network model can reduce training parameters and improve parameter utilization. The final layer of the discriminator network uses the Sigmoid function to complete the normalization. The network model is trained using the weighted sum of loss function, using a score to evaluate the ability of defogging, saving the model and parameters with higher scores during the training process, and finally using the network model with the highest score and parameters to perform the dehazing test on the outdoor real image. The training set uses real fog-free images and synthetic foggy images both indoors and outdoors. The test results show that the proposed model has improved subjective visual effects, image detail and color compared with the traditional dehazing method.