Active Contour Loss based on sobel filter
| 6 | |
| 7 | |
| 8 | class ACLoss(nn.Module): |
| 9 | """ |
| 10 | Active Contour Loss |
| 11 | based on sobel filter |
| 12 | """ |
| 13 | |
| 14 | def __init__(self, miu=1.0, classes=3): |
| 15 | super(ACLoss, self).__init__() |
| 16 | |
| 17 | self.miu = miu |
| 18 | self.classes = classes |
| 19 | sobel_x = np.array([[-1, 0, 1], [-2, 0, 2], [-1, 0, 1]]) |
| 20 | sobel_y = np.array([[-1, -2, -1], [0, 0, 0], [1, 2, 1]]) |
| 21 | |
| 22 | self.sobel_x = nn.Parameter(torch.from_numpy(sobel_x).float().expand(self.classes, 1, 3, 3), |
| 23 | requires_grad=False) |
| 24 | self.sobel_y = nn.Parameter(torch.from_numpy(sobel_y).float().expand(self.classes, 1, 3, 3), |
| 25 | requires_grad=False) |
| 26 | |
| 27 | self.diff_x = nn.Conv2d(self.classes, self.classes, groups=self.classes, kernel_size=3, stride=1, padding=1, |
| 28 | bias=False) |
| 29 | self.diff_x.weight = self.sobel_x |
| 30 | self.diff_y = nn.Conv2d(self.classes, self.classes, groups=self.classes, kernel_size=3, stride=1, padding=1, |
| 31 | bias=False) |
| 32 | self.diff_y.weight = self.sobel_y |
| 33 | |
| 34 | def forward(self, predication, label): |
| 35 | grd_x = self.diff_x(predication) |
| 36 | grd_y = self.diff_y(predication) |
| 37 | |
| 38 | # length |
| 39 | length = torch.sum( |
| 40 | torch.abs(torch.sqrt(grd_x ** 2 + grd_y ** 2 + 1e-8))) |
| 41 | length = (length - length.min()) / (length.max() - length.min() + 1e-8) |
| 42 | length = torch.sum(length) |
| 43 | |
| 44 | # region |
| 45 | label = label.float() |
| 46 | c_in = torch.ones_like(predication) |
| 47 | c_out = torch.zeros_like(predication) |
| 48 | region_in = torch.abs(torch.sum(predication * ((label - c_in) ** 2))) |
| 49 | region_out = torch.abs( |
| 50 | torch.sum((1 - predication) * ((label - c_out) ** 2))) |
| 51 | region = self.miu * region_in + region_out |
| 52 | |
| 53 | return region + length |
| 54 | |
| 55 | |
| 56 | class ACLossV2(nn.Module): |