Image_manipulation_detection-master
所属分类:图形图像处理
开发工具:Python
文件大小:12363KB
下载次数:10
上传日期:2019-01-11 15:26:33
上 传 者:
毛竹meng
说明: 基于RGB流和噪声流的图像篡改检测,该工程提供了完整的实现技术。
(Based on image tampering detection of RGB stream and noise stream, the project provides a complete implementation technology.)
文件列表:
LICENSE (1057, 2018-08-19)
Learning Rich Features for Image Manipulation Dection.pdf (2569043, 2018-08-19)
_image (0, 2018-08-19)
_image\11.png (180034, 2018-08-19)
_image\results.png (1019375, 2018-08-19)
_pdf (0, 2018-08-19)
_pdf\WIFS_2015_splicebuster.pdf (4618161, 2018-08-19)
data (0, 2018-08-19)
data\VOCDevkit2007 (0, 2018-08-19)
data\VOCDevkit2007\VOC2007 (0, 2018-08-19)
data\coco (0, 2018-08-19)
data\coco\LuaAPI (0, 2018-08-19)
data\coco\LuaAPI\CocoApi.lua (10987, 2018-08-19)
data\coco\LuaAPI\MaskApi.lua (10439, 2018-08-19)
data\coco\LuaAPI\cocoDemo.lua (814, 2018-08-19)
data\coco\LuaAPI\env.lua (447, 2018-08-19)
data\coco\LuaAPI\init.lua (511, 2018-08-19)
data\coco\LuaAPI\rocks (0, 2018-08-19)
data\coco\LuaAPI\rocks\coco-scm-1.rockspec (857, 2018-08-19)
data\coco\MatlabAPI (0, 2018-08-19)
data\coco\MatlabAPI\CocoApi.m (14357, 2018-08-19)
data\coco\MatlabAPI\CocoEval.m (22798, 2018-08-19)
data\coco\MatlabAPI\CocoUtils.m (16940, 2018-08-19)
data\coco\MatlabAPI\MaskApi.m (5095, 2018-08-19)
data\coco\MatlabAPI\cocoDemo.m (1234, 2018-08-19)
data\coco\MatlabAPI\evalDemo.m (1852, 2018-08-19)
data\coco\MatlabAPI\gason.m (2448, 2018-08-19)
data\coco\MatlabAPI\private (0, 2018-08-19)
data\coco\MatlabAPI\private\gasonMex.cpp (9457, 2018-08-19)
data\coco\MatlabAPI\private\gasonMex.mexa64 (38020, 2018-08-19)
data\coco\MatlabAPI\private\gasonMex.mexmaci64 (41452, 2018-08-19)
data\coco\MatlabAPI\private\getPrmDflt.m (2994, 2018-08-19)
data\coco\MatlabAPI\private\maskApiMex.c (5678, 2018-08-19)
data\coco\MatlabAPI\private\maskApiMex.mexa64 (21419, 2018-08-19)
data\coco\MatlabAPI\private\maskApiMex.mexmaci64 (23228, 2018-08-19)
data\coco\PythonAPI (0, 2018-08-19)
... ...
# Image_manipulation_detection
Paper: CVPR2018, [Learning Rich Features for Image Manipulation Detection](https://arxiv.org/pdf/1805.04953.pdf)
Code based on [Faster-RCNN](https://github.com/dBeker/Faster-RCNN-TensorFlow-Python3.5)
This is a rough implementation of the paper. Since I do not have a titan gpu, I made some modifications on the algorithm, but you can easily change them back if you want the exact setting from the paper.
![](_image/11.png)
# Environment
Python 3.6
TensorFlow 1.8.0
# Setup
- Download vgg16 pre-trained weights from [here](http://download.tensorflow.org/models/vgg_16_2016_08_28.tar.gz)
- save to /data/imagenet_weights/vgg16.ckpt
- Two-stream neural network model: [lib/nets/vgg16.py](lib/nets/vgg16.py)
- noise stream's weights are randomly initialized
- for accurate prediction, please pre-train noise stream's vgg weights on `ImageNet` and overwrite the trainable setting of noise stream after `SRM` conv layer
- Bounding boxes are predicted by both streams.
- In the paper, `RGB stream` alone predicts bbox more accurately, so you may wanna change that as well (also defined in vgg16.py)
- Use `main_create_training_set.py` to create training set from `PASCAL VOC` dataset.
- The generated dataset will follow the `pascal voc` style, which is also required by `train.py`
- `Tensorboard` file will be save at `/default`
- Weights will be save to `/default/DIY_detaset/default`
# Note
The code requires a large memory GPU. If you do not have a 6G+ GPU, please reduce the number of noise stream conv layers for training.
# Demo results
Dataset size: 10000, epoch: 3
![](_image/results.png)
# Finally
I will update this repo a few weeks later after I installed the new GPU
近期下载者:
相关文件:
收藏者: