# Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. # Refer to # https://github.com/GeoZcx/A-deeply-supervised-image-fusion-network-for-change-detection-in-remote-sensing-images . import paddle import paddle.nn as nn import paddle.nn.functional as F from paddle.vision.models import vgg16 from .layers import Conv1x1, make_norm, ChannelAttention, SpatialAttention class DSIFN(nn.Layer): """ The DSIFN implementation based on PaddlePaddle. The original article refers to C. Zhang, et al., "A deeply supervised image fusion network for change detection in high resolution bi-temporal remote sensing images" (https://www.sciencedirect.com/science/article/pii/S0924271620301532). Note that in this implementation, there is a flexible number of target classes. Args: num_classes (int): Number of target classes. use_dropout (bool, optional): A bool value that indicates whether to use dropout layers. When the model is trained on a relatively small dataset, the dropout layers help prevent overfitting. Default: False. """ def __init__(self, num_classes, use_dropout=False): super(DSIFN, self).__init__() self.encoder1 = self.encoder2 = VGG16FeaturePicker() self.sa1 = SpatialAttention() self.sa2 = SpatialAttention() self.sa3 = SpatialAttention() self.sa4 = SpatialAttention() self.sa5 = SpatialAttention() self.ca1 = ChannelAttention(in_ch=1024) self.bn_ca1 = make_norm(1024) self.o1_conv1 = conv2d_bn(1024, 512, use_dropout) self.o1_conv2 = conv2d_bn(512, 512, use_dropout) self.bn_sa1 = make_norm(512) self.o1_conv3 = Conv1x1(512, num_classes) self.trans_conv1 = nn.Conv2DTranspose(512, 512, kernel_size=2, stride=2) self.ca2 = ChannelAttention(in_ch=1536) self.bn_ca2 = make_norm(1536) self.o2_conv1 = conv2d_bn(1536, 512, use_dropout) self.o2_conv2 = conv2d_bn(512, 256, use_dropout) self.o2_conv3 = conv2d_bn(256, 256, use_dropout) self.bn_sa2 = make_norm(256) self.o2_conv4 = Conv1x1(256, num_classes) self.trans_conv2 = nn.Conv2DTranspose(256, 256, kernel_size=2, stride=2) self.ca3 = ChannelAttention(in_ch=768) self.o3_conv1 = conv2d_bn(768, 256, use_dropout) self.o3_conv2 = conv2d_bn(256, 128, use_dropout) self.o3_conv3 = conv2d_bn(128, 128, use_dropout) self.bn_sa3 = make_norm(128) self.o3_conv4 = Conv1x1(128, num_classes) self.trans_conv3 = nn.Conv2DTranspose(128, 128, kernel_size=2, stride=2) self.ca4 = ChannelAttention(in_ch=384) self.o4_conv1 = conv2d_bn(384, 128, use_dropout) self.o4_conv2 = conv2d_bn(128, 64, use_dropout) self.o4_conv3 = conv2d_bn(64, 64, use_dropout) self.bn_sa4 = make_norm(64) self.o4_conv4 = Conv1x1(64, num_classes) self.trans_conv4 = nn.Conv2DTranspose(64, 64, kernel_size=2, stride=2) self.ca5 = ChannelAttention(in_ch=192) self.o5_conv1 = conv2d_bn(192, 64, use_dropout) self.o5_conv2 = conv2d_bn(64, 32, use_dropout) self.o5_conv3 = conv2d_bn(32, 16, use_dropout) self.bn_sa5 = make_norm(16) self.o5_conv4 = Conv1x1(16, num_classes) self.init_weight() def forward(self, t1, t2): # Extract bi-temporal features. with paddle.no_grad(): self.encoder1.eval(), self.encoder2.eval() t1_feats = self.encoder1(t1) t2_feats = self.encoder2(t2) t1_f_l3, t1_f_l8, t1_f_l15, t1_f_l22, t1_f_l29 = t1_feats t2_f_l3, t2_f_l8, t2_f_l15, t2_f_l22, t2_f_l29, = t2_feats aux_x = [] # Multi-level decoding x = paddle.concat([t1_f_l29, t2_f_l29], axis=1) x = self.o1_conv1(x) x = self.o1_conv2(x) x = self.sa1(x) * x x = self.bn_sa1(x) if self.training: aux_x.append(x) x = self.trans_conv1(x) x = paddle.concat([x, t1_f_l22, t2_f_l22], axis=1) x = self.ca2(x) * x x = self.o2_conv1(x) x = self.o2_conv2(x) x = self.o2_conv3(x) x = self.sa2(x) * x x = self.bn_sa2(x) if self.training: aux_x.append(x) x = self.trans_conv2(x) x = paddle.concat([x, t1_f_l15, t2_f_l15], axis=1) x = self.ca3(x) * x x = self.o3_conv1(x) x = self.o3_conv2(x) x = self.o3_conv3(x) x = self.sa3(x) * x x = self.bn_sa3(x) if self.training: aux_x.append(x) x = self.trans_conv3(x) x = paddle.concat([x, t1_f_l8, t2_f_l8], axis=1) x = self.ca4(x) * x x = self.o4_conv1(x) x = self.o4_conv2(x) x = self.o4_conv3(x) x = self.sa4(x) * x x = self.bn_sa4(x) if self.training: aux_x.append(x) x = self.trans_conv4(x) x = paddle.concat([x, t1_f_l3, t2_f_l3], axis=1) x = self.ca5(x) * x x = self.o5_conv1(x) x = self.o5_conv2(x) x = self.o5_conv3(x) x = self.sa5(x) * x x = self.bn_sa5(x) out5 = self.o5_conv4(x) if not self.training: return [out5] else: size = paddle.shape(t1)[2:] out1 = F.interpolate( self.o1_conv3(aux_x[0]), size=size, mode='bilinear', align_corners=True) out2 = F.interpolate( self.o2_conv4(aux_x[1]), size=size, mode='bilinear', align_corners=True) out3 = F.interpolate( self.o3_conv4(aux_x[2]), size=size, mode='bilinear', align_corners=True) out4 = F.interpolate( self.o4_conv4(aux_x[3]), size=size, mode='bilinear', align_corners=True) return [out5, out4, out3, out2, out1] def init_weight(self): # Do nothing pass class VGG16FeaturePicker(nn.Layer): def __init__(self, indices=(3, 8, 15, 22, 29)): super(VGG16FeaturePicker, self).__init__() features = list(vgg16(pretrained=True).features)[:30] self.features = nn.LayerList(features) self.features.eval() self.indices = set(indices) def forward(self, x): picked_feats = [] for idx, model in enumerate(self.features): x = model(x) if idx in self.indices: picked_feats.append(x) return picked_feats def conv2d_bn(in_ch, out_ch, with_dropout=True): lst = [ nn.Conv2D( in_ch, out_ch, kernel_size=3, stride=1, padding=1), nn.PReLU(), make_norm(out_ch), ] if with_dropout: lst.append(nn.Dropout(p=0.6)) return nn.Sequential(*lst)