每日Attention学习6——Context Aggregation Module

模块出处

$link$ $code$ $IJCAI 22$ Boundary-Guided Camouflaged Object Detection

模块名称

Context Aggregation Module (CAM)

模块作用

增大感受野，全局特征提取

模块结构

模块代码

python 复制代码

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


class ConvBNR(nn.Module):
    def __init__(self, inplanes, planes, kernel_size=3, stride=1, dilation=1, bias=False):
        super(ConvBNR, self).__init__()

        self.block = nn.Sequential(
            nn.Conv2d(inplanes, planes, kernel_size, stride=stride, padding=dilation, dilation=dilation, bias=bias),
            nn.BatchNorm2d(planes),
            nn.ReLU(inplace=True)
        )

    def forward(self, x):
        return self.block(x)


class Conv1x1(nn.Module):
    def __init__(self, inplanes, planes):
        super(Conv1x1, self).__init__()
        self.conv = nn.Conv2d(inplanes, planes, 1)
        self.bn = nn.BatchNorm2d(planes)
        self.relu = nn.ReLU(inplace=True)

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

        return x
    

class CAM(nn.Module):
    def __init__(self, hchannel, channel):
        super(CAM, self).__init__()
        self.conv1_1 = Conv1x1(hchannel + channel, channel)
        self.conv3_1 = ConvBNR(channel // 4, channel // 4, 3)
        self.dconv5_1 = ConvBNR(channel // 4, channel // 4, 3, dilation=2)
        self.dconv7_1 = ConvBNR(channel // 4, channel // 4, 3, dilation=3)
        self.dconv9_1 = ConvBNR(channel // 4, channel // 4, 3, dilation=4)
        self.conv1_2 = Conv1x1(channel, channel)
        self.conv3_3 = ConvBNR(channel, channel, 3)

    def forward(self, lf, hf):
        if lf.size()[2:] != hf.size()[2:]:
            hf = F.interpolate(hf, size=lf.size()[2:], mode='bilinear', align_corners=False)
        x = torch.cat((lf, hf), dim=1)
        x = self.conv1_1(x)
        xc = torch.chunk(x, 4, dim=1)
        x0 = self.conv3_1(xc[0] + xc[1])
        x1 = self.dconv5_1(xc[1] + x0 + xc[2])
        x2 = self.dconv7_1(xc[2] + x1 + xc[3])
        x3 = self.dconv9_1(xc[3] + x2)
        xx = self.conv1_2(torch.cat((x0, x1, x2, x3), dim=1))
        x = self.conv3_3(x + xx)

        return x

    
if __name__ == '__main__':
    x1 = torch.randn([3, 256, 16, 16])
    x2 = torch.randn([3, 512, 8, 8])
    cam = CAM(hchannel=512, channel=256)
    out = cam(x1, x2)
    print(out.shape)  # 3, 256, 16, 16

原文表述

为了将多层次的融合特征整合到伪装物体预测中，我们设计了一个上下文聚合模块（CAM）来挖掘上下文语义，以增强物体检测，如图5所示。不同于BBSNet中的全局上下文模块不考虑各分支之间的语义关联，CAM考虑到跨尺度交互作用以增强特征表示。