研究目的
To propose an end-to-end learned method based on generative adversarial networks (GANs) for tackling the deblurring problem for remote sensing images without any prior assumptions for the blurs.
研究成果
The proposed GAN-based method for remote sensing image deblurring avoids the inaccuracy brought by prior blur assumptions and achieves competitive results compared to other state-of-the-art methods. Future work includes investigating more advanced upsampling methods to minimize checkerboard artifacts.
研究不足
Checkerboard artifacts still exist in some deblurred images, produced by the deconv/transposed convolution operation.
1：Experimental Design and Method Selection:
The proposed method uses CGAN (conditional generative adversarial networks) as the base model with a multi-component loss function, including Wasserstein GAN with gradient penalty and perceptual loss.
2：Sample Selection and Data Sources:
1096 images scraped from Google Maps, with 548 images used for training and the rest for testing.
3：List of Experimental Equipment and Materials:
Implemented using Tensorflow deep learning framework, trained on a single Geforce GTX 1060 GPU with 6GB memory.
4：Experimental Procedures and Operational Workflow:
The generator network takes the blurred image as input and produces the estimate of the sharp image. The critic network takes the restored and sharp images and outputs a distance between them.
5：Data Analysis Methods:
The method was evaluated on standard metrics like peak signal-to-noise ratio (PSNR) and the structural similarity measure (SSIM).
独家科研数据包，助您复现前沿成果，加速创新突破
获取完整内容
