Paper

• Published 2019
• ICCV workshop paper

Introduction

• GAN을 이용한 이미지의 빗물 제거

• Single Image Raindrop Removal/ Spatio-Temporal Raindrop Removal

  ⇒ Matching Dataset이 필요(Rain image/ Non Rainy Image)

Raindrop Generation

• Dataset Augmentation
• Game에서 많이 사용되는 Refraction과 depth of field 기술 사용

→ 더 공부 예정...

Single Image Raindrop Removal

• Raindrop Location Estimator/ Remover 구성

1. Raindrop Location Estimator

→ Encoder - Decoder Architecture

conv layer (stride : 2) - residual block - transposed conv layer * 2

→ Input : Rainy Image, output : addictive location map

→ Loss

Add adversarial loss(Discriminator Loss)

2. Raindrop Removal

→ Architecture

Raindrop Location Estimator와 같은 구조(Input 만 다름)

→ Input : Rainy Image, addictive location map을 channel wise concat

Output : Clean Image(C')

→ Loss : VGG Perceptual Loss + Adversarial Loss