• bababa123
    了解作者
  • C/C++
    开发工具
  • 2.1MB
    文件大小
  • zip
    文件格式
  • 0
    收藏次数
  • 1 积分
    下载积分
  • 8
    下载次数
  • 2018-07-10 16:53
    上传日期
深度学习的超分辨率处理算法,多图像超分处理,亲测可用
2016_super_resolution-master.zip
内容介绍
# 2016_super_resolution [Image Super-Resolution Using Deep Convolutional Networks](https://arxiv.org/abs/1501.00092) ICCV2015 I re-implement this paper and includes my train and test code in this repository. This code uses MIT License. ### Note that: Thanks for @star4s. I fixed some bugs in the network training code and made the code more clear to use. (2017/4/29) # Training data I random selected about 60,000 pic from 2014 ILSVR2014_train (only academic) You can download from [GoogleDriver](https://drive.google.com/open?id=0B0VOCNYh8HeRZmk3SHdrdlcxbXc) or [BaiduYun](https://pan.baidu.com/s/1c0TvFyw) # Result This code get the better performance than 'bicubic' for enlarging a 2x pic. It can be trained and tested now. original pic -> super resolution pic (trained by matconvnet) ![](https://github.com/layumi/2016_super_resolution/blob/master/3_bicubic.jpg) ![](https://github.com/layumi/2016_super_resolution/blob/master/3_srnet.jpg) ![](https://github.com/layumi/2016_super_resolution/blob/master/4_bicubic.jpg) ![](https://github.com/layumi/2016_super_resolution/blob/master/4_srnet.jpg) # How to train & test 1.You may compile matconvnet first by running `gpu_compile.m` (you need to change some setting in it) For more compile information, you can learn it from www.vlfeat.org/matconvnet/install/#compiling 2.run `testSRnet_result.m` for test result. 3.If you want to train it by yourself, you may download my data and use `prepare_ur_data.m` to produce `imdb.mat` which include every picture path. 4.Use `train_SRnet.m` to have fun~ (I also provide a verson for gray-scale images. But the improvement is limited. You can learn more from `train_SRnet_gray.m` and `testSRnet_gray.m`) # Small Tricks 1.I fix the scale factor 2(than 2+2*rand). It seems to be easy for net to learn more information. 2.How to initial net? (You can learn more from `/matlab/+dagnn/@DagNN/initParam.m`) In this work, the initial weight is important! # Citation We greatly appreciate it if you can cite the website in your publications: ``` @misc{2016_super_resolution, title = {{2016_super_resolution}}, howpublished = "\url{https://github.com/layumi/2016_super_resolution}", } ```
评论
    相关推荐
    • 深度学习 Transformers
      使用Transformers的端到端视频实例分割
    • 深度学习.zip
      关于深度学习的一些英文论文,都是我的导师给我推荐的,特别经典,非常值得学习
    • 深度学习入门精选论文
      从起源到细化分类,英文20篇左右,中文十篇,都是经典必看的
    • image_super_resolution:深度学习的图像超分辨率
      一个旨在提高Image Super Resolution深度学习原理/框架/工具的项目。 参考 该项目从以下项目中复制了很多代码和构想: 网络 该模型基于RRDN和GAN。 损失和超参数 HR与SR之间产生损失L1损失 GAN损失,SR的鉴别器...
    • 深度学习论文集二
      论文包含有: 2.1_Model; 2.2_Optimization; 2.3_Unsupervised_Learning___Deep_Generative_Model 2.4_RNN___Sequence-to-Sequence_Model 2.5_Neural_Turing_Machine 2.6_Deep_Reinforcement_Learning ...
    • 深度学习论文汇总
      经典论文汇总,近几年深度学习在CV领域代表作,有需要的同学可以一起打包下载学习
    • 深度学习论文合集
      深度学习方向总结的一些经典的文章,本人也是开始刷论文。慢慢看
    • 深度学习语义分割论文
      深度学习语义分割经典论文,主要为2014到2017年的论文
    • 深度学习目标检测论文
      基于深度学习的目标检测论文包括faster RCNN,SSD, YOLO,Mask RCNN等等,以及各种网络论文:LeNet,AlexNet、vgg,ResNet等等
    • 深度学习超分辨率代码
      使用对抗生成网络,来实现超分辨率。使用imagenet的部分数据。摒弃了用均方误差和信噪比来优化参数。