论文：https://arxiv.org/pdf/2405.06228
代码：https://github.com/nizhenliang/CGRSeg/blob/main/models/decode_heads/rcm.py
Rectangular Self-Calibration Module (RCM) 可以捕捉轴向全局上下文，旨在使模型聚焦前景。
其由矩形自校准注意力 和 MLP 组成，主要结构如下:
矩形自校准注意力 矩形自校准注意力可表述如下:
其中
然后将注意力特征与输入特征进行融合，通过一个卷积层提取输入特征的局部细节，然后与得到的注意力权重相乘
完整代码实现如下：
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 class RCA (nn.Module): def __init__ (self, inp, kernel_size=1 , ratio=1 , band_kernel_size=11 , dw_size=(1 , 1 0 , 0 1 , square_kernel_size=2 , relu=True ): super (RCA, self ).__init__() self .dwconv_hw = nn.Conv2d(inp, inp, square_kernel_size, padding=square_kernel_size // 2 , groups=inp) self .pool_h = nn.AdaptiveAvgPool2d((None , 1 )) self .pool_w = nn.AdaptiveAvgPool2d((1 , None )) gc = inp // ratio self .excite = nn.Sequential( nn.Conv2d(inp, gc, kernel_size=(1 , band_kernel_size), padding=(0 , band_kernel_size // 2 ), groups=gc), nn.BatchNorm2d(gc), nn.ReLU(inplace=True ), nn.Conv2d(gc, inp, kernel_size=(band_kernel_size, 1 ), padding=(band_kernel_size // 2 , 0 ), groups=gc), nn.Sigmoid() ) def sge (self, x ): x_h = self .pool_h(x) x_w = self .pool_w(x) x_gather = x_h + x_w ge = self .excite(x_gather) return ge def forward (self, x ): loc = self .dwconv_hw(x) att = self .sge(x) out = att * loc return out
首先通过一个卷积层提取空间局部特征 loc，输出的特征图形状为 [B, C, H, W]
然后通过分别沿着 H 和 W 两个方向进行平均池化 ，获取这两个方向的全局特征，并将这两个特征相加，得到新特征 x_gather
两个方向的全局特征形状分别为 [B, C, H, 1] 和 [B, C, 1, W]，相加后会自动进行广播，得到的特征图形状为 [B, C, H, W]
然后通过一个 gc，然后进行批归一化和 ReLU 激活函数
随后再使用一个 [0, 1] 之间得到注意力权重
最后将权重与提取的特征 loc 相乘，完成特征融合，重新加权输入特征，增强特征图的表达能力
MLP 通过两个
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 class ConvMlp (nn.Module): """ 使用 1x1 卷积保持空间维度的 MLP """ def __init__ ( self, in_features, hidden_features=None , out_features=None , act_layer=nn.ReLU, norm_layer=None , bias=True , drop=0. ): super ().__init__() out_features = out_features or in_features hidden_features = hidden_features or in_features bias = to_2tuple(bias) self .fc1 = nn.Conv2d(in_features, hidden_features, kernel_size=1 , bias=bias[0 ]) self .norm = norm_layer(hidden_features) if norm_layer else nn.Identity() self .act = act_layer() self .drop = nn.Dropout(drop) self .fc2 = nn.Conv2d(hidden_features, out_features, kernel_size=1 , bias=bias[1 ]) def forward (self, x ): x = self .fc1(x) x = self .norm(x) x = self .act(x) x = self .drop(x) x = self .fc2(x) return x
RCM 的完整实现 在矩形自校准注意力之后添加了批量归一化和 MLP 来完善特征，最后通过残差连接进一步加强特征重用，其可以描述为
$$\mathbf{F}{\mathrm{out}}=\rho(\xi {\mathrm{F}}(\mathrm{x},\xi_{\mathrm{C}}(\mathrm{H}{\mathrm{P}}(\mathrm{x})\oplus\mathrm{V} {\mathrm{P}}(\mathrm{x}))))+\mathrm{x}$$
其中
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 class RCM (nn.Module): """ MetaNeXtBlock 块 参数: dim (int): 输入通道数. drop_path (float): 随机深度率。默认: 0.0 ls_init_value (float): 层级比例初始化值。默认: 1e-6. """ def __init__ ( self, dim, token_mixer=RCA, norm_layer=nn.BatchNorm2d, mlp_layer=ConvMlp, mlp_ratio=2 , act_layer=nn.GELU, ls_init_value=1e-6 , drop_path=0. , dw_size=11 , square_kernel_size=3 , ratio=1 , ): super ().__init__() self .token_mixer = token_mixer(dim, band_kernel_size=dw_size, square_kernel_size=square_kernel_size, ratio=ratio) self .norm = norm_layer(dim) self .mlp = mlp_layer(dim, int (mlp_ratio * dim), act_layer=act_layer) self .gamma = nn.Parameter(ls_init_value * torch.ones(dim)) if ls_init_value else None self .drop_path = DropPath(drop_path) if drop_path > 0. else nn.Identity() def forward (self, x ): shortcut = x x = self .token_mixer(x) x = self .norm(x) x = self .mlp(x) if self .gamma is not None : x = x.mul(self .gamma.reshape(1 , -1 , 1 , 1 )) x = self .drop_path(x) + shortcut return x
