论文阅读笔记(一):Learning Dual Convolutional Neural Networks for Low-Level Vision

Learning Dual Convolutional Neural Networks for Low-Level Vision

论文作者:Jinshan Pan1 Sifei Liu2 Deqing Sun2 Jiawei Zhang3 Yang Liu4 Jimmy Ren5
Zechao Li1 Jinhui Tang1 Huchuan Lu4 Yu-Wing Tai6 Ming-Hsuan Yang7
1Nanjing University of Science and Technology 2NVIDIA 3City University of Hong Kong
4Dalian University of Technology 5SenseTime Research 6Tencent Youtu Lab 7UC Merced

论文开源代码:https://sites.google.com/site/jspanhomepage/dualcnn

针对低层视觉问题提出了一个dualCNN结构,这种架构适用于超分辨、边缘保持性滤波、derain、dehazing一系列的任务。

DualCNN由两部分组成:结构恢复部分(Net-S)和细节恢复部分(Net-D)

Net-S架构:

layer parameters stride padding
1 64 conv 9*9 + relu  1 4
2 32 conv 1*1 + relu 1 0
3 1 conv 5*5 + relu 1 2

Net-D架构:

由20个卷积层组成:64 conv 3*3 + relu

论文阅读笔记(一):Learning Dual Convolutional Neural Networks for Low-Level Vision_第1张图片

训练过程设置:

batch_size = 64

learning_rate = 0.0001

optimize : SGD

训练数据:使用NYU depth dataset 生成训练图像,图像块设置成:32*32

损失函数设置:

论文阅读笔记(一):Learning Dual Convolutional Neural Networks for Low-Level Vision_第2张图片

论文结果比较:

论文阅读笔记(一):Learning Dual Convolutional Neural Networks for Low-Level Vision_第3张图片

你可能感兴趣的:(论文阅读笔记)