/super-resolution

Primary LanguagePython

Single Image Super-Resolution with GAN on HCI Data

Please see Examples.

All results, including PSNR and SSIM, are available in the folder test_out_images.

Based on this paper.

For Cell Images

Please refer to this link for similar experiments for cell images.

The results for cell images are better. One explanation is that cell images are of 1-channel, and the structure of cell images is simple.

Dataset

Train&Val: VOC2012

Test: HCI lightfield Dataset

Examples

Below are examples from papillon, statue, and stillLife.

The GAN-based model produced sharper super-resolution images compared with bicubic, especially when the scaling factor is large. However, when scaling factor is 8, the performance is not as good as in cell images, but GAN is still better than bicubic.

up scale by 2 (full image)

low-res bicubic GAN original
low-res bicubic GAN original
low-res bicubic GAN original
low-res bicubic GAN original

up scale by 2 (details)

bicubic GAN original
bicubic GAN original
bicubic GAN original
bicubic GAN original

up scale by 4 (full image)

low-res bicubic GAN original
low-res bicubic GAN original
low-res bicubic GAN original
low-res bicubic GAN original

up scale by 4 (details)

bicubic GAN original
bicubic GAN original
bicubic GAN original
bicubic GAN original

up scale by 8 (full image)

low-res bicubic GAN original
low-res bicubic GAN original
low-res bicubic GAN original
low-res bicubic GAN original

up scale by 8 (details)

bicubic GAN original
bicubic GAN original
bicubic GAN original
bicubic GAN original

Examples of Cell Images

Link