代码是结合代尔夫特理工大学的deep learning这门课的大作业来讲的。整体上一个分类的框架，但是用了srm噪声提取器，这个后来被RGBN作为双流fasterrcnn中的噪声流支路引用，srm在篡改检测中是比较常见的人工设计的算子。这篇文章的代码我跑过，本身还是有效果的，其中它里面数据预训练的分patch操作也在后续中被引进，但说实话，这种简单的网络设计其实跑不过不做任何设计的分类模型，比如res2net。篡改检测常用的基本都是自然场景数据集，比如CASIA1/2，中科院出的数据集，还有BHSig60,COVERAGE,NC16等，这些数据集的篡改手段包括了resize，压缩啊这些操作，其实和常规的文档类的数据集的ps还是有差别的，我们常说自然场景和文档篡改的差别挺大的，文档主要还是以ps这种操作为主。
上面这张图是网络设计，核心是第一组蓝色的cnn操作，这里面融合例如srm滤波器。srm滤波器最早出自于Rich models for steganalysis of digital images中，是Steganalysis Rich Model的缩写，富隐写分析模型，篡改操作会带来十分尖锐的边缘，尤其是拼接操作，会被高通滤波器很明显的体现出来。文章在cnn的结构里对第一层卷积的权重使用30个在srm中九三残差图的基本高通滤波器进行初始化，这些基础的滤波器对应了7个srm残差类别，分布如下：
| 个数 | 滤波器类别 |
|---|---|
| 8 | 1st |
| 4 | 2nd |
| 8 | 3rd |
| 1 | SQUARE3x3 |
| 4 | EDGE3x3 |
| 1 | SQUARE5x5 |
| 4 | EDGE3x3 |
https://github.com/kPsarakis/Image-Forgery-Detection-CNNfrom typing import Dict
import numpy as np
from torch import Tensor, stack
def get_filters():
"""
Function that return the required high pass SRM filters for the first convolutional layer of our implementation
:return: A pytorch Tensor containing the 30x3x5x5 filter tensor with type
[number_of_filters, input_channels, height, width]
"""
filters: Dict[str, Tensor] = {}
# 1st Order
filters["1O1"] = Tensor(np.array([[0, 0, 0, 0, 0], [0, 0, 0, 0, 0], [0, 0, -1, 1, 0], [0, 0, 0, 0, 0], [0, 0, 0, 0, 0]]))
filters["1O2"] = Tensor(np.rot90(filters["1O1"]).copy())
filters["1O3"] = Tensor(np.rot90(filters["1O2"]).copy())
filters["1O4"] = Tensor(np.rot90(filters["1O3"]).copy())
filters["1O5"] = Tensor(np.array([[0, 0, 0, 0, 0], [0, 0, 0, 1, 0], [0, 0, -1, 0, 0], [0, 0, 0, 0, 0], [0, 0, 0, 0, 0]]))
filters["1O6"] = Tensor(np.rot90(filters["1O5"]).copy())
filters["1O7"] = Tensor(np.rot90(filters["1O6"]).copy())
filters["1O8"] = Tensor(np.rot90(filters["1O7"]).copy())
# 2nd Order
filters["2O1"] = Tensor(np.array([[0, 0, 0, 0, 0], [0, 0, 0, 0, 0], [0, 1, -2, 1, 0], [0, 0, 0, 0, 0], [0, 0, 0, 0, 0]]))
filters["2O2"] = Tensor(np.rot90(filters["2O1"]).copy())
filters["2O3"] = Tensor(np.array([[0, 0, 0, 0, 0], [0, 1, 0, 0, 0], [0, 0, -2, 0, 0], [0, 0, 0, 1, 0], [0, 0, 0, 0, 0]]))
filters["2O4"] = Tensor(np.rot90(filters["2O3"]).copy())
# 3rd Order
filters["3O1"] = Tensor(np.array([[0, 0, 0, 0, 0], [0, 0, 1, 0, 0], [0, 1, -3, 1, 0], [0, 0, 0, 0, 0], [0, 0, 0, 0, 0]]))
filters["3O2"] = Tensor(np.rot90(filters["3O1"]).copy())
filters["3O3"] = Tensor(np.rot90(filters["3O2"]).copy())
filters["3O4"] = Tensor(np.rot90(filters["3O3"]).copy())
filters["3O5"] = Tensor(np.array([[0, 0, 0, 0, 0], [0, 1, 0, 1, 0], [0, 0, -3, 0, 0], [0, 1, 0, 0, 0], [0, 0, 0, 0, 0]]))
filters["3O6"] = Tensor(np.rot90(filters["3O5"]).copy())
filters["3O7"] = Tensor(np.rot90(filters["3O6"]).copy())
filters["3O8"] = Tensor(np.rot90(filters["3O7"]).copy())
# 3x3 SQUARE
filters["3x3S"] = Tensor(np.array([[0, 0, 0, 0, 0], [0, -1, 2, -1, 0], [0, 2, -4, 2, 0], [0, -1, 2, -1, 0], [0, 0, 0, 0, 0]]))
# 3x3 EDGE
filters["3x3E1"] = Tensor(np.array([[0, 0, 0, 0, 0], [0, -1, 2, -1, 0], [0, 2, -4, 2, 0], [0, 0, 0, 0, 0], [0, 0, 0, 0, 0]]))
filters["3x3E2"] = Tensor(np.rot90(filters["3x3E1"]).copy())
filters["3x3E3"] = Tensor(np.rot90(filters["3x3E2"]).copy())
filters["3x3E4"] = Tensor(np.rot90(filters["3x3E3"]).copy())
# 5X5 EDGE
filters["5x5E1"] = Tensor(np.array([[-1, 2, -2, 2, -1], [2, -6, 8, -6, 2], [-2, 8, -12, 8, -2], [0, 0, 0, 0, 0], [0, 0, 0, 0, 0]]))
filters["5x5E2"] = Tensor(np.rot90(filters["5x5E1"]).copy())
filters["5x5E3"] = Tensor(np.rot90(filters["5x5E2"]).copy())
filters["5x5E4"] = Tensor(np.rot90(filters["5x5E3"]).copy())
# 5x5 SQUARE
filters["5x5S"] = Tensor(np.array([[-1, 2, -2, 2, -1], [2, -6, 8, -6, 2], [-2, 8, -12, 8, -2], [2, -6, 8, -6, 2], [-1, 2, -2, 2, -1]]))
return vectorize_filters(filters)
def vectorize_filters(filters: dict):
"""
Function that takes as input the 30x5x5 different SRM high pass filters and creates the 30x3x5x5 tensor with the
following permutations