YOLOv8轻量化改进——Coordinate Attention注意力机制

现在针对YOLOv8的架构改进越来越多，今天尝试引入了Coordinate Attention注意力机制以改进对小目标物体的检测效率。

yolov8的下载和安装参考我这篇博客：

基于SeaShips数据集的yolov8训练教程_seaships处理成yolov8-CSDN博客

首先我们可以去官网找到CA注意力机制模块的相关文件：

官网链接：GitHub - houqb/CoordAttention: Code for our CVPR2021 paper coordinate attention

接下来我们就可以去下载好的yolov8配置文件中进行修改：

首先是modules.py文件，路径如下：

在最下方添加CA注意力机制类的定义：

class CA(nn.Module):
    def __init__(self, inp, oup, reduction=32):
        super(CA, self).__init__()
        torch.use_deterministic_algorithms(True, warn_only=True)  # 允许非确定性操作但继续运行
        self.pool_h = nn.AdaptiveAvgPool2d((None, 1))
        self.pool_w = nn.AdaptiveAvgPool2d((1, None))
        mip = max(8, inp // reduction)

        self.conv1 = nn.Conv2d(inp, mip, kernel_size=1, stride=1, padding=0)
        self.bn1 = nn.BatchNorm2d(mip)
        self.act = h_swish()
        
        self.conv_h = nn.Conv2d(mip, oup, kernel_size=1, stride=1, padding=0)
        self.conv_w = nn.Conv2d(mip, oup, kernel_size=1, stride=1, padding=0)
        

    def forward(self, x):
        identity = x
        
        n,c,h,w = x.size()
        x_h = self.pool_h(x)
        x_w = self.pool_w(x).permute(0, 1, 3, 2)

        y = torch.cat([x_h, x_w], dim=2)
        y = self.conv1(y)
        y = self.bn1(y)
        y = self.act(y) 
        
        x_h, x_w = torch.split(y, [h, w], dim=2)
        x_w = x_w.permute(0, 1, 3, 2)

        a_h = self.conv_h(x_h).sigmoid()
        a_w = self.conv_w(x_w).sigmoid()

        out = identity * a_w * a_h

        return out

结果如下：

接着去tasks.py文件下：

最上方注册CA模块：

用以下代码：

from ultralytics.nn.modules import (C1, C2, C3, C3TR, SPP, SPPF, Bottleneck, BottleneckCSP, C2f, C3Ghost, C3x, Classify,
                                    Concat, Conv, ConvTranspose, Detect, DWConv, DWConvTranspose2d, Ensemble, Focus,
                                    GhostBottleneck, GhostConv, Segment,CA,DualConv)

然后找到parse_model函数：

可以Crtl+F全局搜索一下比较快，找到其中一行：

if m in (Classify, Conv, ConvTranspose, GhostConv, Bottleneck, GhostBottleneck, SPP, SPPF, DWConv, Focus,
                 BottleneckCSP, C1, C2, C2f, C3, C3TR, C3Ghost, nn.ConvTranspose2d, DWConvTranspose2d, C3x):

在最后C3x后面加入CA即可。

最后再找到你的yaml配置文件：

修改backbone和检测头head，根据你需要添加CA机制的地方来引入CA模块，注意前后输入输出的一致性，以及添加层之后的层序号变动，这些都是需要考虑的，这里给出我的配置文件代码供参考：

# Parameters
nc: 6  # number of classes
depth_multiple: 1.00  # scales module repeats
width_multiple: 1.00  # scales convolution channels

# YOLOv8.0l backbone-only CA
backbone:
  # [from, repeats, module, args]
  - [-1, 1, Conv, [64, 3, 2]]   # 0-P1/2 [3,64,320,320]
  - [-1, 1, Conv, [128, 3, 2]]  # 1-P2/4 [64,128,160,160]
  - [-1, 3, C2f, [128, True]]   # 2      [128,128,160,160]
  - [-1, 1, CA, [128]]           # 3      [128,128,160,160]
  - [-1, 1, Conv, [256, 3, 2]]  # 4-P3/8 [128,256,80,80]
  - [-1, 6, C2f, [256, True]]   # 5      [256,256,80,80]
  - [-1, 1, CA, [256]]           # 6      [256,256,80,80]
  - [-1, 1, Conv, [512, 3, 2]]  # 7-P4/16 [256,512,40,40]
  - [-1, 6, C2f, [512, True]]   # 8       [512,512,40,40]
  - [-1, 1, CA, [512]]           # 9       [512,512,40,40]
  - [-1, 1, Conv, [512, 3, 2]]  # 10-P5/32 [512,512,20,20]
  - [-1, 3, C2f, [512, True]]   # 11       [512,512,20,20]
  - [-1, 1, CA, [512]]           # 12       [512,512,20,20]
  - [-1, 1, SPPF, [512, 5]]     # 13      [512,512,20,20]

# YOLOv8.0l head
head:
  - [-1, 1, nn.Upsample, [None, 2, 'nearest']]  # 14 [512,512,40,40]
  - [[-1, 8], 1, Concat, [1]]                   # 15 拼接 [512,512,40,40] + [512,512,40,40] → [1024,512,40,40]
  - [-1, 3, C2f, [512]]                        # 16 [1024,512,40,40] → [512,512,40,40]
  - [-1, 1, nn.Upsample, [None, 2, 'nearest']]  # 17 [512,512,80,80]
  - [[-1, 5], 1, Concat, [1]]                   # 18 拼接 [512,512,80,80] + [256,256,80,80] → [768,768,80,80]
  - [-1, 3, C2f, [256]]                        # 19 [768,256,80,80]
  - [-1, 1, Conv, [256, 3, 2]]                  # 20 [256,256,40,40]
  - [[-1, 16], 1, Concat, [1]]                  # 21 拼接 [256,256,40,40] + [512,512,40,40] → [768,768,40,40]
  - [-1, 3, C2f, [512]]                        # 22 [768,512,40,40]
  - [-1, 1, Conv, [512, 3, 2]]                  # 23 [512,512,20,20]
  - [[-1, 13], 1, Concat, [1]]                  # 24 拼接 [512,512,20,20] + [512,512,20,20] → [1024,1024,20,20]
  - [-1, 3, C2f, [512]]                        # 25 [1024,512,20,20]
  - [[19, 22, 25], 1, Detect, [nc]]            # 26 Detect(P3, P4, P5)

这样就全部改进完毕了，接下来运行命令就可以看结果了。