BiseNet实现遥感影像地物分类

遥感地物分类通过对遥感图像中的地物进行准确识别和分类，为资源管理、环境保护、城市规划、灾害监测等领域提供重要信息，有助于实现精细化管理和科学决策，提升社会治理和经济发展水平。深度学习遥感地物分类在提高分类精度、自动化程度、处理大规模数据、普适性以及推动遥感技术创新和发展等方面都具有重要的意义。本文将利用深度学习BiseNet实现遥感地物分类。

数据集

本文使用的数据集为WHDLD数据集^[1](Wuhan dense labeling dataset)。WHDLD数据集包括4940张高分辨率遥感影像，包含6种土地覆盖类型，影像尺寸均被裁剪至256×256像素。下面是一些数据集示例。

BiSeNet

BiseNet^[2]（Bilateral Segmentation Network）是一种用于图像分割的深度学习网络。它具有双边分割的特点，可以同时处理空间信息和上下文信息，从而实现高效、准确的图像分割。

具体来说，BiseNet由两个分支组成：空间路径（spatial path）和上下文路径（context path）。其中，空间路径具有较小的感受野，可以捕获丰富的空间信息并生成高分辨率的特征图；而上下文路径则具有较大的感受野，可以捕获更多的上下文信息并生成低分辨率的特征图。这两个路径通过一个特征融合模块进行融合，从而生成既包含丰富空间信息又包含上下文信息的分割结果。

在BiseNet中，还有一些关键的技术和设计，如轻量级模型设计、注意力机制、特征融合等，这些技术和设计可以进一步提升网络的性能和效率。

网络复现

resnet18

import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.utils.model_zoo as modelzoo

resnet18_url = 'https://download.pytorch.org/models/resnet18-5c106cde.pth'


from torch.nn import BatchNorm2d


def conv3x3(in_planes, out_planes, stride=1):
    """3x3 convolution with padding"""
    return nn.Conv2d(in_planes, out_planes, kernel_size=3, stride=stride,
                     padding=1, bias=False)


class BasicBlock(nn.Module):
    def __init__(self, in_chan, out_chan, stride=1):
        super(BasicBlock, self).__init__()
        self.conv1 = conv3x3(in_chan, out_chan, stride)
        self.bn1 = BatchNorm2d(out_chan)
        self.conv2 = conv3x3(out_chan, out_chan)
        self.bn2 = BatchNorm2d(out_chan)
        self.relu = nn.ReLU(inplace=True)
        self.downsample = None
        if in_chan != out_chan or stride != 1:
            self.downsample = nn.Sequential(
                nn.Conv2d(in_chan, out_chan,
                          kernel_size=1, stride=stride, bias=False),
                BatchNorm2d(out_chan),
                )

    def forward(self, x):
        residual = self.conv1(x)
        residual = self.bn1(residual)
        residual = self.relu(residual)
        residual = self.conv2(residual)
        residual = self.bn2(residual)

        shortcut = x
        if self.downsample is not None:
            shortcut = self.downsample(x)

        out = shortcut + residual
        out = self.relu(out)
        return out


def create_layer_basic(in_chan, out_chan, bnum, stride=1):
    layers = [BasicBlock(in_chan, out_chan, stride=stride)]
    for i in range(bnum-1):
        layers.append(BasicBlock(out_chan, out_chan, stride=1))
    return nn.Sequential(*layers)


class Resnet18(nn.Module):
    def __init__(self):
        super(Resnet18, self).__init__()
        self.conv1 = nn.Conv2d(3, 64, kernel_size=7, stride=2, padding=3,
                               bias=False)
        self.bn1 = BatchNorm2d(64)
        self.relu = nn.ReLU(inplace=True)
        self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)
        self.layer1 = create_layer_basic(64, 64, bnum=2, stride=1)
        self.layer2 = create_layer_basic(64, 128, bnum=2, stride=2)
        self.layer3 = create_layer_basic(128, 256, bnum=2, stride=2)
        self.layer4 = create_layer_basic(256, 512, bnum=2, stride=2)
        self.init_weight()

    def forward(self, x):
        x = self.conv1(x)
        x = self.bn1(x)
        x = self.relu(x)
        x = self.maxpool(x)

        x = self.layer1(x)
        feat8 = self.layer2(x) # 1/8
        feat16 = self.layer3(feat8) # 1/16
        feat32 = self.layer4(feat16) # 1/32
        return feat8, feat16, feat32

    def init_weight(self):
        state_dict = modelzoo.load_url(resnet18_url)
        self_state_dict = self.state_dict()
        for k, v in state_dict.items():
            if 'fc' in k: continue
            self_state_dict.update({k: v})
        self.load_state_dict(self_state_dict)

    def get_params(self):
        wd_params, nowd_params = [], []
        for name, module in self.named_modules():
            if isinstance(module, (nn.Linear, nn.Conv2d)):
                wd_params.append(module.weight)
                if not module.bias is None:
                    nowd_params.append(module.bias)
            elif isinstance(module, nn.modules.batchnorm._BatchNorm):
                nowd_params += list(module.parameters())
        return wd_params, nowd_params

BiSeNet

import torch
import torch.nn as nn
import torch.nn.functional as F
import torchvision

from .resnet import Resnet18

from torch.nn import BatchNorm2d


class ConvBNReLU(nn.Module):

    def __init__(self, in_chan, out_chan, ks=3, stride=1, padding=1, *args, **kwargs):
        super(ConvBNReLU, self).__init__()
        self.conv = nn.Conv2d(in_chan,
                out_chan,
                kernel_size = ks,
                stride = stride,
                padding = padding,
                bias = False)
        self.bn = BatchNorm2d(out_chan)
        self.relu = nn.ReLU(inplace=True)
        self.init_weight()

    def forward(self, x):
        x = self.conv(x)
        x = self.bn(x)
        x = self.relu(x)
        return x

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


class UpSample(nn.Module):

    def __init__(self, n_chan, factor=2):
        super(UpSample, self).__init__()
        out_chan = n_chan * factor * factor
        self.proj = nn.Conv2d(n_chan, out_chan, 1, 1, 0)
        self.up = nn.PixelShuffle(factor)
        self.init_weight()

    def forward(self, x):
        feat = self.proj(x)
        feat = self.up(feat)
        return feat

    def init_weight(self):
        nn.init.xavier_normal_(self.proj.weight, gain=1.)


class BiSeNetOutput(nn.Module):

    def __init__(self, in_chan, mid_chan, n_classes, up_factor=32, *args, **kwargs):
        super(BiSeNetOutput, self).__init__()
        self.up_factor = up_factor
        out_chan = n_classes
        self.conv = ConvBNReLU(in_chan, mid_chan, ks=3, stride=1, padding=1)
        self.conv_out = nn.Conv2d(mid_chan, out_chan, kernel_size=1, bias=True)
        self.up = nn.Upsample(scale_factor=up_factor,
                mode='bilinear', align_corners=False)
        self.init_weight()

    def forward(self, x):
        x = self.conv(x)
        x = self.conv_out(x)
        x = self.up(x)
        return x

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

    def get_params(self):
        wd_params, nowd_params = [], []
        for name, module in self.named_modules():
            if isinstance(module, (nn.Linear, nn.Conv2d)):
                wd_params.append(module.weight)
                if not module.bias is None:
                    nowd_params.append(module.bias)
            elif isinstance(module, nn.modules.batchnorm._BatchNorm):
                nowd_params += list(module.parameters())
        return wd_params, nowd_params


class AttentionRefinementModule(nn.Module):
    def __init__(self, in_chan, out_chan, *args, **kwargs):
        super(AttentionRefinementModule, self).__init__()
        self.conv = ConvBNReLU(in_chan, out_chan, ks=3, stride=1, padding=1)
        self.conv_atten = nn.Conv2d(out_chan, out_chan, kernel_size= 1, bias=False)
        self.bn_atten = BatchNorm2d(out_chan)
        #  self.sigmoid_atten = nn.Sigmoid()
        self.init_weight()

    def forward(self, x):
        feat = self.conv(x)
        atten = torch.mean(feat, dim=(2, 3), keepdim=True)
        atten = self.conv_atten(atten)
        atten = self.bn_atten(atten)
        #  atten = self.sigmoid_atten(atten)
        atten = atten.sigmoid()
        out = torch.mul(feat, atten)
        return out

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


class ContextPath(nn.Module):
    def __init__(self, *args, **kwargs):
        super(ContextPath, self).__init__()
        self.resnet = Resnet18()
        self.arm16 = AttentionRefinementModule(256, 128)
        self.arm32 = AttentionRefinementModule(512, 128)
        self.conv_head32 = ConvBNReLU(128, 128, ks=3, stride=1, padding=1)
        self.conv_head16 = ConvBNReLU(128, 128, ks=3, stride=1, padding=1)
        self.conv_avg = ConvBNReLU(512, 128, ks=1, stride=1, padding=0)
        self.up32 = nn.Upsample(scale_factor=2.)
        self.up16 = nn.Upsample(scale_factor=2.)

        self.init_weight()

    def forward(self, x):
        feat8, feat16, feat32 = self.resnet(x)

        avg = torch.mean(feat32, dim=(2, 3), keepdim=True)
        avg = self.conv_avg(avg)

        feat32_arm = self.arm32(feat32)
        feat32_sum = feat32_arm + avg
        feat32_up = self.up32(feat32_sum)
        feat32_up = self.conv_head32(feat32_up)

        feat16_arm = self.arm16(feat16)
        feat16_sum = feat16_arm + feat32_up
        feat16_up = self.up16(feat16_sum)
        feat16_up = self.conv_head16(feat16_up)

        return feat16_up, feat32_up # x8, x16

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

    def get_params(self):
        wd_params, nowd_params = [], []
        for name, module in self.named_modules():
            if isinstance(module, (nn.Linear, nn.Conv2d)):
                wd_params.append(module.weight)
                if not module.bias is None:
                    nowd_params.append(module.bias)
            elif isinstance(module, nn.modules.batchnorm._BatchNorm):
                nowd_params += list(module.parameters())
        return wd_params, nowd_params


class SpatialPath(nn.Module):
    def __init__(self, *args, **kwargs):
        super(SpatialPath, self).__init__()
        self.conv1 = ConvBNReLU(3, 64, ks=7, stride=2, padding=3)
        self.conv2 = ConvBNReLU(64, 64, ks=3, stride=2, padding=1)
        self.conv3 = ConvBNReLU(64, 64, ks=3, stride=2, padding=1)
        self.conv_out = ConvBNReLU(64, 128, ks=1, stride=1, padding=0)
        self.init_weight()

    def forward(self, x):
        feat = self.conv1(x)
        feat = self.conv2(feat)
        feat = self.conv3(feat)
        feat = self.conv_out(feat)
        return feat

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

    def get_params(self):
        wd_params, nowd_params = [], []
        for name, module in self.named_modules():
            if isinstance(module, nn.Linear) or isinstance(module, nn.Conv2d):
                wd_params.append(module.weight)
                if not module.bias is None:
                    nowd_params.append(module.bias)
            elif isinstance(module, nn.modules.batchnorm._BatchNorm):
                nowd_params += list(module.parameters())
        return wd_params, nowd_params


class FeatureFusionModule(nn.Module):
    def __init__(self, in_chan, out_chan, *args, **kwargs):
        super(FeatureFusionModule, self).__init__()
        self.convblk = ConvBNReLU(in_chan, out_chan, ks=1, stride=1, padding=0)
        ## use conv-bn instead of 2 layer mlp, so that tensorrt 7.2.3.4 can work for fp16
        self.conv = nn.Conv2d(out_chan,
                out_chan,
                kernel_size = 1,
                stride = 1,
                padding = 0,
                bias = False)
        self.bn = nn.BatchNorm2d(out_chan)
        #  self.conv1 = nn.Conv2d(out_chan,
        #          out_chan//4,
        #          kernel_size = 1,
        #          stride = 1,
        #          padding = 0,
        #          bias = False)
        #  self.conv2 = nn.Conv2d(out_chan//4,
        #          out_chan,
        #          kernel_size = 1,
        #          stride = 1,
        #          padding = 0,
        #          bias = False)
        #  self.relu = nn.ReLU(inplace=True)
        self.init_weight()

    def forward(self, fsp, fcp):
        fcat = torch.cat([fsp, fcp], dim=1)
        feat = self.convblk(fcat)
        atten = torch.mean(feat, dim=(2, 3), keepdim=True)
        atten = self.conv(atten)
        atten = self.bn(atten)
        #  atten = self.conv1(atten)
        #  atten = self.relu(atten)
        #  atten = self.conv2(atten)
        atten = atten.sigmoid()
        feat_atten = torch.mul(feat, atten)
        feat_out = feat_atten + feat
        return feat_out

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

    def get_params(self):
        wd_params, nowd_params = [], []
        for name, module in self.named_modules():
            if isinstance(module, (nn.Linear, nn.Conv2d)):
                wd_params.append(module.weight)
                if not module.bias is None:
                    nowd_params.append(module.bias)
            elif isinstance(module, nn.modules.batchnorm._BatchNorm):
                nowd_params += list(module.parameters())
        return wd_params, nowd_params


class BiSeNetV1(nn.Module):

    def __init__(self, n_classes, aux_mode='train', *args, **kwargs):
        super(BiSeNetV1, self).__init__()
        self.cp = ContextPath()
        self.sp = SpatialPath()
        self.ffm = FeatureFusionModule(256, 256)
        self.conv_out = BiSeNetOutput(256, 256, n_classes, up_factor=8)
        self.aux_mode = aux_mode
        if self.aux_mode == 'train':
            self.conv_out16 = BiSeNetOutput(128, 64, n_classes, up_factor=8)
            self.conv_out32 = BiSeNetOutput(128, 64, n_classes, up_factor=16)
        self.init_weight()

    def forward(self, x):
        H, W = x.size()[2:]
        feat_cp8, feat_cp16 = self.cp(x)
        feat_sp = self.sp(x)
        feat_fuse = self.ffm(feat_sp, feat_cp8)

        feat_out = self.conv_out(feat_fuse)
        if self.aux_mode == 'train':
            feat_out16 = self.conv_out16(feat_cp8)
            feat_out32 = self.conv_out32(feat_cp16)
            return feat_out, feat_out16, feat_out32
        elif self.aux_mode == 'eval':
            return feat_out,
        elif self.aux_mode == 'pred':
            feat_out = feat_out.argmax(dim=1)
            return feat_out
        else:
            raise NotImplementedError

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

    def get_params(self):
        wd_params, nowd_params, lr_mul_wd_params, lr_mul_nowd_params = [], [], [], []
        for name, child in self.named_children():
            child_wd_params, child_nowd_params = child.get_params()
            if isinstance(child, (FeatureFusionModule, BiSeNetOutput)):
                lr_mul_wd_params += child_wd_params
                lr_mul_nowd_params += child_nowd_params
            else:
                wd_params += child_wd_params
                nowd_params += child_nowd_params
        return wd_params, nowd_params, lr_mul_wd_params, lr_mul_nowd_params

训练过程精度变化

测试精度

结果展示

总结

今天的分享到此结束，感兴趣的点点关注，后续将分享更多案例。

参考资料

[1]

WHDLD: https://sites.google.com/view/zhouwx/dataset#h.p_hQS2jYeaFpV0

[2]

BiSeNet: https://arxiv.org/abs/1808.00897

本文由 mdnice 多平台发布