unetformer实现遥感建筑物提取

最新推荐文章于 2024-07-05 10:45:32 发布

DataAssassin

最新推荐文章于 2024-07-05 10:45:32 发布

阅读量1.1k

点赞数 18

文章标签：深度学习

本文链接：https://blog.csdn.net/weixin_44095417/article/details/137764220

版权

前言

本期将分享「UNetFormer网络」，论文地址https://www.sciencedirect.com/science/article/abs/pii/S0924271622001654?via%3Dihub。源码地址https://github.com/WangLibo1995/GeoSeg

数据集

本文选取的是WHU-Building-DataSets。数据集^[1]包含了从新西兰基督城的航空图像中提取的超过220,000个独立建筑，图像被分割成了8189个512×512像素的片，其中包含了训练集（130,500个建筑），验证集（14,500个建筑）和测试集（42,000个建筑）。

UNetFormer

此次我们使用的模型为UNetFormer。UNetFormer提出了一种基于transformer的解码器，一种高效的全局-局部注意机制global-local Transformer block (GLTB) ，用于实时城市场景分割。。

网络结构


import torch
import torch.nn as nn
import torch.nn.functional as F
from einops import rearrange, repeat

from timm.models.layers import DropPath, to_2tuple, trunc_normal_
import timm


class ConvBNReLU(nn.Sequential):
    def __init__(self, in_channels, out_channels, kernel_size=3, dilation=1, stride=1, norm_layer=nn.BatchNorm2d, bias=False):
        super(ConvBNReLU, self).__init__(
            nn.Conv2d(in_channels, out_channels, kernel_size=kernel_size, bias=bias,
                      dilation=dilation, stride=stride, padding=((stride - 1) + dilation * (kernel_size - 1)) // 2),
            norm_layer(out_channels),
            nn.ReLU6()
        )


class ConvBN(nn.Sequential):
    def __init__(self, in_channels, out_channels, kernel_size=3, dilation=1, stride=1, norm_layer=nn.BatchNorm2d, bias=False):
        super(ConvBN, self).__init__(
            nn.Conv2d(in_channels, out_channels, kernel_size=kernel_size, bias=bias,
                      dilation=dilation, stride=stride, padding=((stride - 1) + dilation * (kernel_size - 1)) // 2),
            norm_layer(out_channels)
        )


class Conv(nn.Sequential):
    def __init__(self, in_channels, out_channels, kernel_size=3, dilation=1, stride=1, bias=False):
        super(Conv, self).__init__(
            nn.Conv2d(in_channels, out_channels, kernel_size=kernel_size, bias=bias,
                      dilation=dilation, stride=stride, padding=((stride - 1) + dilation * (kernel_size - 1)) // 2)
        )


class SeparableConvBNReLU(nn.Sequential):
    def __init__(self, in_channels, out_channels, kernel_size=3, stride=1, dilation=1,
                 norm_layer=nn.BatchNorm2d):
        super(SeparableConvBNReLU, self).__init__(
            nn.Conv2d(in_channels, in_channels, kernel_size, stride=stride, dilation=dilation,
                      padding=((stride - 1) + dilation * (kernel_size - 1)) // 2,
                      groups=in_channels, bias=False),
            norm_layer(out_channels),
            nn.Conv2d(in_channels, out_channels, kernel_size=1, bias=False),
            nn.ReLU6()
        )


class SeparableConvBN(nn.Sequential):
    def __init__(self, in_channels, out_channels, kernel_size=3, stride=1, dilation=1,
                 norm_layer=nn.BatchNorm2d):
        super(SeparableConvBN, self).__init__(
            nn.Conv2d(in_channels, in_channels, kernel_size, stride=stride, dilation=dilation,
                      padding=((stride - 1) + dilation * (kernel_size - 1)) // 2,
                      groups=in_channels, bias=False),
            norm_layer(out_channels),
            nn.Conv2d(in_channels, out_channels, kernel_size=1, bias=False)
        )


class SeparableConv(nn.Sequential):
    def __init__(self, in_channels, out_channels, kernel_size=3, stride=1, dilation=1):
        super(SeparableConv, self).__init__(
            nn.Conv2d(in_channels, in_channels, kernel_size, stride=stride, dilation=dilation,
                      padding=((stride - 1) + dilation * (kernel_size - 1)) // 2,
                      groups=in_channels, bias=False),
            nn.Conv2d(in_channels, out_channels, kernel_size=1, bias=False)
        )


class Mlp(nn.Module):
    def __init__(self, in_features, hidden_features=None, out_features=None, act_layer=nn.ReLU6, drop=0.):
        super().__init__()
        out_features = out_features or in_features
        hidden_features = hidden_features or in_features
        self.fc1 = nn.Conv2d(in_features, hidden_features, 1, 1, 0, bias=True)
        self.act = act_layer()
        self.fc2 = nn.Conv2d(hidden_features, out_features, 1, 1, 0, bias=True)
        self.drop = nn.Dropout(drop, inplace=True)

    def forward(self, x):
        x = self.fc1(x)
        x = self.act(x)
        x = self.drop(x)
        x = self.fc2(x)
        x = self.drop(x)
        return x


class GlobalLocalAttention(nn.Module):
    def __init__(self,
                 dim=256,
                 num_heads=16,
                 qkv_bias=False,
                 window_size=8,
                 relative_pos_embedding=True
                 ):
        super().__init__()
        self.num_heads = num_heads
        head_dim = dim // self.num_heads
        self.scale = head_dim ** -0.5
        self.ws = window_size

        self.qkv = Conv(dim, 3*dim, kernel_size=1, bias=qkv_bias)
        self.local1 = ConvBN(dim, dim, kernel_size=3)
        self.local2 = ConvBN(dim, dim, kernel_size=1)
        self.proj = SeparableConvBN(dim, dim, kernel_size=window_size)

        self.attn_x = nn.AvgPool2d(kernel_size=(window_size, 1), stride=1,  padding=(window_size//2 - 1, 0))
        self.attn_y = nn.AvgPool2d(kernel_size=(1, window_size), stride=1, padding=(0, window_size//2 - 1))

        self.relative_pos_embedding = relative_pos_embedding

        if self.relative_pos_embedding:
            # define a parameter table of relative position bias
            self.relative_position_bias_table = nn.Parameter(
                torch.zeros((2 * window_size - 1) * (2 * window_size - 1), num_heads))  # 2*Wh-1 * 2*Ww-1, nH

            # get pair-wise relative position index for each token inside the window
            coords_h = torch.arange(self.ws)
            coords_w = torch.arange(self.ws)
            coords = torch.stack(torch.meshgrid([coords_h, coords_w]))  # 2, Wh, Ww
            coords_flatten = torch.flatten(coords, 1)  # 2, Wh*Ww
            relative_coords = coords_flatten[:, :, None] - coords_flatten[:, None, :]  # 2, Wh*Ww, Wh*Ww
            relative_coords = relative_coords.permute(1, 2, 0).contiguous()  # Wh*Ww, Wh*Ww, 2
            relative_coords[:, :, 0] += self.ws - 1  # shift to start from 0
            relative_coords[:, :, 1] += self.ws - 1
            relative_coords[:, :, 0] *= 2 * self.ws - 1
            relative_position_index = relative_coords.sum(-1)  # Wh*Ww, Wh*Ww
            self.register_buffer("relative_position_index", relative_position_index)

            trunc_normal_(self.relative_position_bias_table, std=.02)

    def pad(self, x, ps):
        _, _, H, W = x.size()
        if W % ps != 0:
            x = F.pad(x, (0, ps - W % ps), mode='reflect')
        if H % ps != 0:
            x = F.pad(x, (0, 0, 0, ps - H % ps), mode='reflect')
        return x

    def pad_out(self, x):
        x = F.pad(x, pad=(0, 1, 0, 1), mode='reflect')
        return x

    def forward(self, x):
        B, C, H, W = x.shape

        local = self.local2(x) + self.local1(x)

        x = self.pad(x, self.ws)
        B, C, Hp, Wp = x.shape
        qkv = self.qkv(x)

        q, k, v = rearrange(qkv, 'b (qkv h d) (hh ws1) (ww ws2) -> qkv (b hh ww) h (ws1 ws2) d', h=self.num_heads,
                            d=C//self.num_heads, hh=Hp//self.ws, ww=Wp//self.ws, qkv=3, ws1=self.ws, ws2=self.ws)

        dots = (q @ k.transpose(-2, -1)) * self.scale

        if self.relative_pos_embedding:
            relative_position_bias = self.relative_position_bias_table[self.relative_position_index.view(-1)].view(
                self.ws * self.ws, self.ws * self.ws, -1)  # Wh*Ww,Wh*Ww,nH
            relative_position_bias = relative_position_bias.permute(2, 0, 1).contiguous()  # nH, Wh*Ww, Wh*Ww
            dots += relative_position_bias.unsqueeze(0)

        attn = dots.softmax(dim=-1)
        attn = attn @ v

        attn = rearrange(attn, '(b hh ww) h (ws1 ws2) d -> b (h d) (hh ws1) (ww ws2)', h=self.num_heads,
                         d=C//self.num_heads, hh=Hp//self.ws, ww=Wp//self.ws, ws1=self.ws, ws2=self.ws)

        attn = attn[:, :, :H, :W]

        out = self.attn_x(F.pad(attn, pad=(0, 0, 0, 1), mode='reflect')) + \
              self.attn_y(F.pad(attn, pad=(0, 1, 0, 0), mode='reflect'))

        out = out + local
        out = self.pad_out(out)
        out = self.proj(out)
        # print(out.size())
        out = out[:, :, :H, :W]

        return out


class Block(nn.Module):
    def __init__(self, dim=256, num_heads=16,  mlp_ratio=4., qkv_bias=False, drop=0., attn_drop=0.,
                 drop_path=0., act_layer=nn.ReLU6, norm_layer=nn.BatchNorm2d, window_size=8):
        super().__init__()
        self.norm1 = norm_layer(dim)
        self.attn = GlobalLocalAttention(dim, num_heads=num_heads, qkv_bias=qkv_bias, window_size=window_size)

        self.drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity()
        mlp_hidden_dim = int(dim * mlp_ratio)
        self.mlp = Mlp(in_features=dim, hidden_features=mlp_hidden_dim, out_features=dim, act_layer=act_layer, drop=drop)
        self.norm2 = norm_layer(dim)

    def forward(self, x):

        x = x + self.drop_path(self.attn(self.norm1(x)))
        x = x + self.drop_path(self.mlp(self.norm2(x)))

        return x


class WF(nn.Module):
    def __init__(self, in_channels=128, decode_channels=128, eps=1e-8):
        super(WF, self).__init__()
        self.pre_conv = Conv(in_channels, decode_channels, kernel_size=1)

        self.weights = nn.Parameter(torch.ones(2, dtype=torch.float32), requires_grad=True)
        self.eps = eps
        self.post_conv = ConvBNReLU(decode_channels, decode_channels, kernel_size=3)

    def forward(self, x, res):
        x = F.interpolate(x, scale_factor=2, mode='bilinear', align_corners=False)
        weights = nn.ReLU()(self.weights)
        fuse_weights = weights / (torch.sum(weights, dim=0) + self.eps)
        x = fuse_weights[0] * self.pre_conv(res) + fuse_weights[1] * x
        x = self.post_conv(x)
        return x


class FeatureRefinementHead(nn.Module):
    def __init__(self, in_channels=64, decode_channels=64):
        super().__init__()
        self.pre_conv = Conv(in_channels, decode_channels, kernel_size=1)

        self.weights = nn.Parameter(torch.ones(2, dtype=torch.float32), requires_grad=True)
        self.eps = 1e-8
        self.post_conv = ConvBNReLU(decode_channels, decode_channels, kernel_size=3)

        self.pa = nn.Sequential(nn.Conv2d(decode_channels, decode_channels, kernel_size=3, padding=1, groups=decode_channels),
                                nn.Sigmoid())
        self.ca = nn.Sequential(nn.AdaptiveAvgPool2d(1),
                                Conv(decode_channels, decode_channels//16, kernel_size=1),
                                nn.ReLU6(),
                                Conv(decode_channels//16, decode_channels, kernel_size=1),
                                nn.Sigmoid())

        self.shortcut = ConvBN(decode_channels, decode_channels, kernel_size=1)
        self.proj = SeparableConvBN(decode_channels, decode_channels, kernel_size=3)
        self.act = nn.ReLU6()

    def forward(self, x, res):
        x = F.interpolate(x, scale_factor=2, mode='bilinear', align_corners=False)
        weights = nn.ReLU()(self.weights)
        fuse_weights = weights / (torch.sum(weights, dim=0) + self.eps)
        x = fuse_weights[0] * self.pre_conv(res) + fuse_weights[1] * x
        x = self.post_conv(x)
        shortcut = self.shortcut(x)
        pa = self.pa(x) * x
        ca = self.ca(x) * x
        x = pa + ca
        x = self.proj(x) + shortcut
        x = self.act(x)

        return x


class AuxHead(nn.Module):

    def __init__(self, in_channels=64, num_classes=8):
        super().__init__()
        self.conv = ConvBNReLU(in_channels, in_channels)
        self.drop = nn.Dropout(0.1)
        self.conv_out = Conv(in_channels, num_classes, kernel_size=1)

    def forward(self, x, h, w):
        feat = self.conv(x)
        feat = self.drop(feat)
        feat = self.conv_out(feat)
        feat = F.interpolate(feat, size=(h, w), mode='bilinear', align_corners=False)
        return feat


class Decoder(nn.Module):
    def __init__(self,
                 encoder_channels=(64, 128, 256, 512),
                 decode_channels=64,
                 dropout=0.1,
                 window_size=8,
                 num_classes=6):
        super(Decoder, self).__init__()

        self.pre_conv = ConvBN(encoder_channels[-1], decode_channels, kernel_size=1)
        self.b4 = Block(dim=decode_channels, num_heads=8, window_size=window_size)

        self.b3 = Block(dim=decode_channels, num_heads=8, window_size=window_size)
        self.p3 = WF(encoder_channels[-2], decode_channels)

        self.b2 = Block(dim=decode_channels, num_heads=8, window_size=window_size)
        self.p2 = WF(encoder_channels[-3], decode_channels)

        if self.training:
            self.up4 = nn.UpsamplingBilinear2d(scale_factor=4)
            self.up3 = nn.UpsamplingBilinear2d(scale_factor=2)
            self.aux_head = AuxHead(decode_channels, num_classes)

        self.p1 = FeatureRefinementHead(encoder_channels[-4], decode_channels)

        self.segmentation_head = nn.Sequential(ConvBNReLU(decode_channels, decode_channels),
                                               nn.Dropout2d(p=dropout, inplace=True),
                                               Conv(decode_channels, num_classes, kernel_size=1))
        self.init_weight()

    def forward(self, res1, res2, res3, res4, h, w):
        if self.training:
            x = self.b4(self.pre_conv(res4))
            h4 = self.up4(x)

            x = self.p3(x, res3)
            x = self.b3(x)
            h3 = self.up3(x)

            x = self.p2(x, res2)
            x = self.b2(x)
            h2 = x
            x = self.p1(x, res1)
            x = self.segmentation_head(x)
            x = F.interpolate(x, size=(h, w), mode='bilinear', align_corners=False)

            ah = h4 + h3 + h2
            ah = self.aux_head(ah, h, w)

            return x, ah
        else:
            x = self.b4(self.pre_conv(res4))
            x = self.p3(x, res3)
            x = self.b3(x)

            x = self.p2(x, res2)
            x = self.b2(x)

            x = self.p1(x, res1)

            x = self.segmentation_head(x)
            x = F.interpolate(x, size=(h, w), mode='bilinear', align_corners=False)

            return x

    def init_weight(self):
        for m in self.children():
            if isinstance(m, nn.Conv2d):
                nn.init.kaiming_normal_(m.weight, a=1)
                if m.bias is not None:
                    nn.init.constant_(m.bias, 0)


class UNetFormer(nn.Module):
    def __init__(self,
                 decode_channels=64,
                 dropout=0.1,
                 backbone_name='swsl_resnet18',
                 pretrained=True,
                 window_size=8,
                 num_classes=6
                 ):
        super().__init__()

        self.backbone = timm.create_model(backbone_name, features_only=True, output_stride=32,
                                          out_indices=(1, 2, 3, 4), pretrained=pretrained)
        encoder_channels = self.backbone.feature_info.channels()

        self.decoder = Decoder(encoder_channels, decode_channels, dropout, window_size, num_classes)

    def forward(self, x):
        h, w = x.size()[-2:]
        res1, res2, res3, res4 = self.backbone(x)
        if self.training:
            x, ah = self.decoder(res1, res2, res3, res4, h, w)
            return x, ah
        else:
            x = self.decoder(res1, res2, res3, res4, h, w)
            return x

训练结果

测试结果

结语

「完整代码与训练结果请加入我们的星球。」

「感兴趣的可以加入我们的星球，获取更多数据集、网络复现源码与训练结果的」。

alt 「加入前不要忘了在公众号首页领取优惠券哦！」

往期精彩

参考资料

[1]

WHU-Building-DataSets: https://study.rsgis.whu.edu.cn/pages/download/building_dataset.html

本文由 mdnice 多平台发布

DataAssassin

关注

18
点赞
踩
14

收藏

觉得还不错? 一键收藏
打赏
0
评论
unetformer实现遥感建筑物提取

此次我们使用的模型为UNetFormer。UNetFormer提出了一种基于transformer的解码器，一种高效的全局-局部注意机制global-local Transformer block (GLTB) ，用于实时城市场景分割。包含了从新西兰基督城的航空图像中提取的超过220,000个独立建筑，图像被分割成了8189个512×512像素的片，其中包含了训练集（130,500个建筑），验证集（14,500个建筑）和测试集（42,000个建筑）。
复制链接

扫一扫