zhengyuzhao / perc-adversarial Goto Github PK

您好，我在做关于图像增强方面的工作（颜色，亮度），请问您提出来的这个loss可以用来做类似任务计算损失，比较两张图片的差异性吗，我看到的目前有比较两张图像差异性的有psnr，ssim，lpips，您有做过该方面的对比吗，多谢

Hi,
It's not an issue but maybe you could help.
I'm doing research on adversarial attacks, and I need both conversions (lab=>rgb, rgb=>lab), so I took your code for rgb=>lab, and implemented the other way around using both your code and other sources (skimage package, wikipedia).
for some reason it didn't work for me. I did notice a fixed per-channel shift in the intermediate xyz representations so once I add the following values [1.6836, 1.7713, 1.9287] after the lab=>xyz conversion (and before the xyz=>rgb) - it works perfectly for most pixels, but still not for the darker pixels.
Adding my code here, maybe you'll see what I did wrong...

`def lab2rgb(lab_image, device):
y = (lab_image[:,0,:,:] + 16.) / 116.
x = (lab_image[:,1,:,:] / 500.) + y
z = y - (lab_image[:,2,:,:] / 200.)

assert torch.all(z >= 0)

xyz = torch.stack([x,y,z], dim=1)

mask1 = xyz > 0.2068966
mask1_not = torch.logical_not(mask1)

xyz = (mask1 * xyz) ** 3. + (mask1_not * xyz - 16.0 / 116.) / 7.787

xyz[:,0,:,:] *= 95.0489
xyz[:,1,:,:] *= 100.
xyz[:,2,:,:] *= 108.8840

# here's the shift I added (have no idea why but it works for most pixels)
xyz[:,0,:,:] += 1.6836
xyz[:,1,:,:] += 1.7713
xyz[:,2,:,:] += 1.9287

mt = torch.tensor([[0.4124, 0.3576, 0.1805],
                   [0.2126, 0.7152, 0.0722],
                   [0.0193, 0.1192, 0.9504]]).to(device)
mt_inv = torch.inverse(mt)

rgb = torch.matmul(mt_inv, xyz.permute(1, 0, 2, 3).contiguous().view(3, -1)).view(3, xyz.size(0),
                                                                                   xyz.size(2),
                                                                                   xyz.size(3)).permute(1, 0, 2, 3)
rgb /= 100.
mask2 = rgb > 0.0031308
mask2_not = torch.logical_not(mask2)
rgb = 1.055 * (mask2 * rgb) ** (1/2.4) - 0.055 + mask2_not * rgb * 12.92

return torch.clamp(rgb, 0, 1)`

Thanks in advance :)
Itai

您好，我是中南大学的一名本科生，无意间在csdn上看到了您的这个方法。因为自己是对抗样本领域的门外汉，虽然将您的源码成功运行，但结果是“进程已结束,退出代码0”。
所以想问您运行结果是正常的吗？有成功生成对抗样本了吗？
希望能得到您的答复，谢谢。

请问这部分代码有公布么？暂时没有找到

你好！我想知道关于图1是如何从原图上截取一块并放大的，可以发我一份代码吗?邮箱[email protected]