IEEE 2017 Conference on

Computer Vision and Pattern

Recognition

Introduction

Motivation

Recurrent Joint Rain Detection and Removal Experiment Results

Deep Joint Rain Detection and Removal from a Single ImageWenhan Yang1, Robby T. Tan2,3, Jiashi Feng2, Jiaying Liu1, Zongming Guo1, and Shuicheng Yan4,2

1Institute of Computer Science and Technology, Peking University, Beijing, China2National University of Singapore, 3Yale-NUS College, 4360 AI Institute

IEEE 2017 Conference on Computer

Vision and Pattern Recognition

Contextualized Dilated Network

Visual comparison

Hard case

- The intrinsic overlapping between

rain streaks and texture.

- Complex degradations, e.g. heavy

rain and mist.

- A limited receptive field.

Challenges

- Region-dependent rain models

Rain-streak binary mask

Diversified streaks & accumulation

- Joint rain detection and removal

- Contextualized dilated network

Contributions

Rain Image Formation

1

1s

t

t

O = B + S R A

Region dependent

- Rain localization.

- Allowing a new rain removal pipeline to detect rain regions first,

and then to operate differently.

Mist and heavy rains

Motivation

- A larger receptive field.

- More context information.

ResNet

- Increasingly larger receptive

fields.

Multi-Path Dilation

Convolutions

- Filter at different scales

- P1: 5×5

- P2: 9×9

- P3: 13×13

Rain detection, estimation and removal

F R, F,R S, , ,

F,R S O RS B

Rain Images DSC[1] LP[2] Ours

For more details & codes, scan QR code or navigate

to http://www.icst.pku.edu.cn/struct/Projects/joint_rain_removal.html

Interested in our team STRUCT? Navigate to

1. Y. Li, et al. Rain streak removal using layer

priors. CVPR, 2016.

2. Y. Luo, et al. Removing rain from a single

image via discriminative sparse coding.

ICCV, 2015.

STRUCT @ PKUSpatial and Temporal Restoration, Understanding and Compression

YaleNUSCollege

NUSNational Universityof Singapore

http://www.icst.pku.edu.cn/struct/struct.html