Deraindrop

Visual Attentive GAN Project

Network

• Generator

  

  • Attention map Generation Module

  

  • Autoencoder - Raindrop Free Generation Module

  

  

• Discriminator

---

Result

• PSNR : 34.83 , SSIM : 0.9788
  

• PSNR : 30.88 , SSIM : 0.9588
  

• PSNR : 31.78 , SSIM : 0.9613
  

• PSNR : 33.42 , SSIM : 0.9680
  

• PSNR : 33.59 , SSIM : 0.9697
  

• PSNR : 35.36 , SSIM : 0.9792
  

---

Comparison with Original Attentive GAN

• AttGAN(Input size 720 x 480) : PSNR 26.25 / SSIM 0.9433

• Ours(Input size 360 x 240) : PSNR 36.34 / SSIM 0.9824
  

• AttGAN : PSNR 30.26 / SSIM 0.9482

• Ours : PSNR 35.92 / SSIM 0.9813
  

• AttGAN : PSNR 26.86 / SSIM 0.9576

• Ours : PSNR 26.90 / SSIM 0.9756
  

• AttGAN : PSNR 24.33 / SSIM 0.9482

• Ours : PSNR 36.05 / SSIM 0.9835
  

• AttGAN : PSNR 27.13 / SSIM 0.9548

• Ours : PSNR 34.10 / SSIM 0.9792
  

---

Result of Object Detection YOLO v3

• Raindrop Image & Generative Image

---

References

• LSGAN : https://arxiv.org/pdf/1611.04076.pdf
• SRGAN : https://arxiv.org/pdf/1609.04802.pdf
• RCAN(Image Super-Resolution Using Very Deep Residual Channel Attention Networks) : https://arxiv.org/pdf/1807.02758.pdf
• pix2pix : https://arxiv.org/pdf/1611.07004.pdf
• Neural style transfer : https://arxiv.org/pdf/1508.06576v2.pdf
• Loss Functions for Image Restoration with Neural Networks : https://arxiv.org/pdf/1511.08861.pdf