基于生成对抗网络的无人机图像去模糊方法

UAV Image Deblurring Method Based on Generative Adversarial Network

提出一种基于加权感知损失的生成对抗网络(GAN)用于无人机图像去模糊。实验中采用具有跳跃连接结构的网络作为生成器,并对生成器使用加权感知损失进行约束,在生成器和判别器进行对抗式训练过程中,生成器不断学习并优化模糊图像到对应清晰图像的映射函数。另外,由于PSNR、SSIM图像质量客观评价指标的局限性,提出使用感知损失作为监控网络优化过程和模型选择的评价指标,最后使用感知损失选择的生成器模型对模糊图像进行盲去模糊。实验表明,该方法可快速有效地恢复出细节清晰的图像。

During the process of acquiring UAV image,the UAV can be affected by many factors such as wind,mechanical vibration and rapid relative motion between targets and UAV,so the images can be easily blurred and lose important details.Some CNN-based methods are used to deblur images,however,may still produce smooth images.The intuition is that in classification network,lower layers capture fine details while higher layers capture coarse objects.In this paper,we used network with skip connections as the generator to significantly reduce the inference time.In addition,due to the limitation of image quality assessment metrics such as PSNR and SSIM,we took weighted perceptual loss as an image quality assessment metric for model selection.The experimental result shows that this method can quickly and effectively restore the image with clear details.