语义分割实战——基于EGEUNet神经网络印章分割系统2：含训练测试代码和数据集

第一步：准备数据

印章分割-深度学习图像分割数据集

印章分割数据，可直接应用到一些常用深度学习分割算法中，比如FCN、Unet、SegNet、DeepLabV1、DeepLabV2、DeepLabV3、DeepLabV3+、PSPNet、RefineNet、HRnet、Mask R-CNN、Segformer、DUCK-Net模型等
数据集总共有2000对图片，数据质量非常高，甚至可应用到工业落地的项目中

第二步：搭建模型

本文选择EGEUNet，其网络结构分别如下：

第三步：训练代码

1）损失函数为：dice_loss + focal_loss

2）网络代码：

class EGEUNet(nn.Module):def __init__(self, num_classes=1, input_channels=3, c_list=[8, 16, 24, 32, 48, 64], bridge=True, gt_ds=False):super().__init__()self.bridge = bridgeself.gt_ds = gt_dsself.encoder1 = nn.Sequential(nn.Conv2d(input_channels, c_list[0], 3, stride=1, padding=1),)self.encoder2 = nn.Sequential(nn.Conv2d(c_list[0], c_list[1], 3, stride=1, padding=1),)self.encoder3 = nn.Sequential(nn.Conv2d(c_list[1], c_list[2], 3, stride=1, padding=1),)self.encoder4 = nn.Sequential(Grouped_multi_axis_Hadamard_Product_Attention(c_list[2], c_list[3]),)self.encoder5 = nn.Sequential(Grouped_multi_axis_Hadamard_Product_Attention(c_list[3], c_list[4]),)self.encoder6 = nn.Sequential(Grouped_multi_axis_Hadamard_Product_Attention(c_list[4], c_list[5]),)if bridge:self.GAB1 = group_aggregation_bridge(c_list[1], c_list[0])self.GAB2 = group_aggregation_bridge(c_list[2], c_list[1])self.GAB3 = group_aggregation_bridge(c_list[3], c_list[2])self.GAB4 = group_aggregation_bridge(c_list[4], c_list[3])self.GAB5 = group_aggregation_bridge(c_list[5], c_list[4])print('group_aggregation_bridge was used')if gt_ds:self.gt_conv1 = nn.Sequential(nn.Conv2d(c_list[4], 1, 1))self.gt_conv2 = nn.Sequential(nn.Conv2d(c_list[3], 1, 1))self.gt_conv3 = nn.Sequential(nn.Conv2d(c_list[2], 1, 1))self.gt_conv4 = nn.Sequential(nn.Conv2d(c_list[1], 1, 1))self.gt_conv5 = nn.Sequential(nn.Conv2d(c_list[0], 1, 1))print('gt deep supervision was used')self.decoder1 = nn.Sequential(Grouped_multi_axis_Hadamard_Product_Attention(c_list[5], c_list[4]),)self.decoder2 = nn.Sequential(Grouped_multi_axis_Hadamard_Product_Attention(c_list[4], c_list[3]),)self.decoder3 = nn.Sequential(Grouped_multi_axis_Hadamard_Product_Attention(c_list[3], c_list[2]),)self.decoder4 = nn.Sequential(nn.Conv2d(c_list[2], c_list[1], 3, stride=1, padding=1),)self.decoder5 = nn.Sequential(nn.Conv2d(c_list[1], c_list[0], 3, stride=1, padding=1),)self.ebn1 = nn.GroupNorm(4, c_list[0])self.ebn2 = nn.GroupNorm(4, c_list[1])self.ebn3 = nn.GroupNorm(4, c_list[2])self.ebn4 = nn.GroupNorm(4, c_list[3])self.ebn5 = nn.GroupNorm(4, c_list[4])self.dbn1 = nn.GroupNorm(4, c_list[4])self.dbn2 = nn.GroupNorm(4, c_list[3])self.dbn3 = nn.GroupNorm(4, c_list[2])self.dbn4 = nn.GroupNorm(4, c_list[1])self.dbn5 = nn.GroupNorm(4, c_list[0])self.final = nn.Conv2d(c_list[0], num_classes, kernel_size=1)self.apply(self._init_weights)def _init_weights(self, m):if isinstance(m, nn.Linear):trunc_normal_(m.weight, std=.02)if isinstance(m, nn.Linear) and m.bias is not None:nn.init.constant_(m.bias, 0)elif isinstance(m, nn.Conv1d):n = m.kernel_size[0] * m.out_channelsm.weight.data.normal_(0, math.sqrt(2. / n))elif isinstance(m, nn.Conv2d):fan_out = m.kernel_size[0] * m.kernel_size[1] * m.out_channelsfan_out //= m.groupsm.weight.data.normal_(0, math.sqrt(2.0 / fan_out))if m.bias is not None:m.bias.data.zero_()def forward(self, x):out = F.gelu(F.max_pool2d(self.ebn1(self.encoder1(x)), 2, 2))t1 = out  # b, c0, H/2, W/2out = F.gelu(F.max_pool2d(self.ebn2(self.encoder2(out)), 2, 2))t2 = out  # b, c1, H/4, W/4out = F.gelu(F.max_pool2d(self.ebn3(self.encoder3(out)), 2, 2))t3 = out  # b, c2, H/8, W/8out = F.gelu(F.max_pool2d(self.ebn4(self.encoder4(out)), 2, 2))t4 = out  # b, c3, H/16, W/16out = F.gelu(F.max_pool2d(self.ebn5(self.encoder5(out)), 2, 2))t5 = out  # b, c4, H/32, W/32out = F.gelu(self.encoder6(out))  # b, c5, H/32, W/32t6 = outout5 = F.gelu(self.dbn1(self.decoder1(out)))  # b, c4, H/32, W/32if self.gt_ds:gt_pre5 = self.gt_conv1(out5)t5 = self.GAB5(t6, t5, gt_pre5)gt_pre5 = F.interpolate(gt_pre5, scale_factor=32, mode='bilinear', align_corners=True)else:t5 = self.GAB5(t6, t5)out5 = torch.add(out5, t5)  # b, c4, H/32, W/32out4 = F.gelu(F.interpolate(self.dbn2(self.decoder2(out5)), scale_factor=(2, 2), mode='bilinear',align_corners=True))  # b, c3, H/16, W/16if self.gt_ds:gt_pre4 = self.gt_conv2(out4)t4 = self.GAB4(t5, t4, gt_pre4)gt_pre4 = F.interpolate(gt_pre4, scale_factor=16, mode='bilinear', align_corners=True)else:t4 = self.GAB4(t5, t4)out4 = torch.add(out4, t4)  # b, c3, H/16, W/16out3 = F.gelu(F.interpolate(self.dbn3(self.decoder3(out4)), scale_factor=(2, 2), mode='bilinear',align_corners=True))  # b, c2, H/8, W/8if self.gt_ds:gt_pre3 = self.gt_conv3(out3)t3 = self.GAB3(t4, t3, gt_pre3)gt_pre3 = F.interpolate(gt_pre3, scale_factor=8, mode='bilinear', align_corners=True)else:t3 = self.GAB3(t4, t3)out3 = torch.add(out3, t3)  # b, c2, H/8, W/8out2 = F.gelu(F.interpolate(self.dbn4(self.decoder4(out3)), scale_factor=(2, 2), mode='bilinear',align_corners=True))  # b, c1, H/4, W/4if self.gt_ds:gt_pre2 = self.gt_conv4(out2)t2 = self.GAB2(t3, t2, gt_pre2)gt_pre2 = F.interpolate(gt_pre2, scale_factor=4, mode='bilinear', align_corners=True)else:t2 = self.GAB2(t3, t2)out2 = torch.add(out2, t2)  # b, c1, H/4, W/4out1 = F.gelu(F.interpolate(self.dbn5(self.decoder5(out2)), scale_factor=(2, 2), mode='bilinear',align_corners=True))  # b, c0, H/2, W/2if self.gt_ds:gt_pre1 = self.gt_conv5(out1)t1 = self.GAB1(t2, t1, gt_pre1)gt_pre1 = F.interpolate(gt_pre1, scale_factor=2, mode='bilinear', align_corners=True)else:t1 = self.GAB1(t2, t1)out1 = torch.add(out1, t1)  # b, c0, H/2, W/2out0 = F.interpolate(self.final(out1), scale_factor=(2, 2), mode='bilinear',align_corners=True)  # b, num_class, H, Wif self.gt_ds:return (torch.sigmoid(gt_pre5), torch.sigmoid(gt_pre4), torch.sigmoid(gt_pre3), torch.sigmoid(gt_pre2),torch.sigmoid(gt_pre1)), torch.sigmoid(out0)else:return torch.sigmoid(out0)

第四步：统计一些指标（训练过程中的loss和miou）