基于改进生成对抗网络的红外图像超分辨率重建

由于硬件成本和拍摄条件等限制,很难直接获取高分辨率红外图像。生成对抗网络可以实现红外图像的超分辨率重建,但仍存在训练不稳定,训练时不收敛等不足。针对这些问题,本文使用Wasserstein距离代替KL散度,结合图像间的欧式距离构造新的损失函数,优化原有网络结构和算法流程,使网络更准确地学习低分辨率图像与重建图像的对应特征映射关系,网络训练更加稳定。实验结果表明,重建图像的边缘过渡平缓,目标细节得到有效保证,并获得了更好的客观评价结果。

Infrared image super-resolution based on improved generative adversarial networks

Due to limitations in hardware and shooting conditions,it is hard to obtain high-resolution infrared images. Generative adversarial networks can achieve super-resolution reconstruction of infrared images,but there are still some shortcomings such as insufficient training and no convergence during training. To deal with these problems,this paper proposes an improved method. The Wasserstein distance is used instead of KL divergence,and the new objective function is constructed by combining the Euclidean distance between images. The network can learn the mapping relationship between the low-resolution image and the reconstructed image more accurately,and the network training is more stable. The experimental results show that the edge transition of the reconstructed image is gentle,the target details are effectively guaranteed,and a better objective evaluation result is obtained.