学习资源站

YOLOv11改进-细节创新篇-iAFF迭代注意力特征融合改进C3k2助力yolov11有效涨点

一、本文介绍

本文给大家带来的最新改进机制是 iAFF（迭代注意力特征融合） ，其主要思想是通过改善特征融合过程来提高检测精度。传统的 特征融合方法 如加法或串联简单，未考虑到特定对象的融合适用性。iAFF通过引入多尺度通道注意力模块 (我个人觉得这个改进机制就算融合了注意力机制的求和操作) ，更好地整合不同尺度和语义不一致的特征。 该方法属于细节上的改进 ， 并不影响任何其它的模块，非常适合大家进行融合改进，单独使用也是有一定的涨点效果。

欢迎大家订阅我的专栏一起学习YOLO！

二、iAFF的基本框架原理

​​

官方论文地址： 官方论文地址点击即可跳转

官方代码地址： 官方代码地址点击即可跳转

​​

iAFF的主要思想在于通过更精细的注意力机制来改善特征融合，从而增强 卷积神经网络 。它不仅处理了由于尺度和语义不一致而引起的特征融合问题，还引入了 多尺度 通道注意力模块，提供了一种统一且通用的特征融合方案。此外，iAFF通过迭代注意力特征融合来解决特征图初始整合可能成为的瓶颈。这种方法使得 模型 即使在层数或参数较少的情况下，也能取得到较好的效果。

iAFF的创新点主要包括：

1. 注意力特征融合： 提出了一种新的特征融合方式，利用注意力机制来改善传统的简单特征融合方法（如加和或串联）。

2. 多尺度通道注意力模块： 解决了在不同尺度上融合特征时出现的问题，特别是语义和尺度不一致的特征融合问题。

3. 迭代注意力特征融合（iAFF）： 通过迭代地应用注意力机制来改善特征图的初步整合，克服了初步整合可能成为性能瓶颈的问题。

​​

这张图片是关于所提出的AFF（注意力特征融合）和iAFF（迭代注意力特征融合）的示意图。图中展示了两种结构：

(a) AFF: 展示了一个通过多尺度通道 注意力模块 （MS-CAM）来融合不同特征的基本框架。特征图X和Y通过MS-CAM和其他操作融合，产生输出Z。

(b) iAFF: 与AFF类似，但添加了迭代结构。在这里，输出Z回馈到输入，与X和Y一起再次经过MS-CAM和融合操作，以进一步细化特征融合过程。

(这两种方法都是文章中提出的我仅使用了iAFF也就是更复杂的版本，大家对于AFF有兴趣的可以按照我的该法进行相似添加即可)

三、iAFF的核心代码

使用方式看章节四！ 

  1. import torch
  2. import torch.nn as nn
  6. __all__ = ['C3k2_iAFF']
  9. def autopad(k, p=None, d=1):  # kernel, padding, dilation
  10.     """Pad to 'same' shape outputs."""
  11.     if d > 1:
  12.         k = d * (k - 1) + 1 if isinstance(k, int) else [d * (x - 1) + 1 for x in k]  # actual kernel-size
  13.     if p is None:
  14.         p = k // 2 if isinstance(k, int) else [x // 2 for x in k]  # auto-pad
  15.     return p
  18. class Conv(nn.Module):
  19.     """Standard convolution with args(ch_in, ch_out, kernel, stride, padding, groups, dilation, activation)."""
  20.     default_act = nn.SiLU()  # default activation
  22.     def __init__(self, c1, c2, k=1, s=1, p=None, g=1, d=1, act=True):
  23.         """Initialize Conv layer with given arguments including activation."""
  24.         super().__init__()
  25.         self.conv = nn.Conv2d(c1, c2, k, s, autopad(k, p, d), groups=g, dilation=d, bias=False)
  26.         self.bn = nn.BatchNorm2d(c2)
  27.         self.act = self.default_act if act is True else act if isinstance(act, nn.Module) else nn.Identity()
  29.     def forward(self, x):
  30.         """Apply convolution, batch normalization and activation to input tensor."""
  31.         return self.act(self.bn(self.conv(x)))
  33.     def forward_fuse(self, x):
  34.         """Perform transposed convolution of 2D data."""
  35.         return self.act(self.conv(x))
  38. class iAFF(nn.Module):
  39.     '''
  40.     多特征融合 iAFF
  41.     '''
  43.     def __init__(self, channels=64, r=2):
  44.         super(iAFF, self).__init__()
  45.         inter_channels = int(channels // r)
  47.         # 本地注意力
  48.         self.local_att = nn.Sequential(
  49.             nn.Conv2d(channels, inter_channels, kernel_size=1, stride=1, padding=0),
  50.             nn.BatchNorm2d(inter_channels),
  51.             nn.ReLU(inplace=True),
  52.             nn.Conv2d(inter_channels, channels, kernel_size=1, stride=1, padding=0),
  53.             nn.BatchNorm2d(channels),
  54.         )
  56.         # 全局注意力
  57.         self.global_att = nn.Sequential(
  58.             nn.AdaptiveAvgPool2d(1),
  59.             nn.Conv2d(channels, inter_channels, kernel_size=1, stride=1, padding=0),
  60.             nn.ReLU(inplace=True),
  61.             nn.Conv2d(inter_channels, channels, kernel_size=1, stride=1, padding=0),
  62.         )
  64.         # 第二次本地注意力
  65.         self.local_att2 = nn.Sequential(
  66.             nn.Conv2d(channels, inter_channels, kernel_size=1, stride=1, padding=0),
  67.             nn.BatchNorm2d(inter_channels),
  68.             nn.ReLU(inplace=True),
  69.             nn.Conv2d(inter_channels, channels, kernel_size=1, stride=1, padding=0),
  70.             nn.BatchNorm2d(channels),
  71.         )
  72.         # 第二次全局注意力
  73.         self.global_att2 = nn.Sequential(
  74.             nn.AdaptiveAvgPool2d(1),
  75.             nn.Conv2d(channels, inter_channels, kernel_size=1, stride=1, padding=0),
  76.             nn.BatchNorm2d(inter_channels),
  77.             nn.ReLU(inplace=True),
  78.             nn.Conv2d(inter_channels, channels, kernel_size=1, stride=1, padding=0),
  79.             nn.BatchNorm2d(channels),
  80.         )
  82.         self.sigmoid = nn.Sigmoid()
  84.     def forward(self, x, residual):
  86.         xa = x + residual
  87.         xl = self.local_att(xa)
  88.         xg = self.global_att(xa)
  89.         xlg = xl + xg
  90.         wei = self.sigmoid(xlg)
  91.         xi = x * wei + residual * (1 - wei)
  93.         xl2 = self.local_att2(xi)
  94.         xg2 = self.global_att(xi)
  95.         xlg2 = xl2 + xg2
  96.         wei2 = self.sigmoid(xlg2)
  97.         xo = x * wei2 + residual * (1 - wei2)
  98.         return xo
  101. class AFF(nn.Module):
  102.     '''
  103.     多特征融合 AFF
  104.     '''
  106.     def __init__(self, channels=64, r=4):
  107.         super(AFF, self).__init__()
  108.         inter_channels = int(channels // r)
  110.         self.local_att = nn.Sequential(
  111.             nn.Conv2d(channels, inter_channels, kernel_size=1, stride=1, padding=0),
  112.             nn.BatchNorm2d(inter_channels),
  113.             nn.ReLU(inplace=True),
  114.             nn.Conv2d(inter_channels, channels, kernel_size=1, stride=1, padding=0),
  115.             nn.BatchNorm2d(channels),
  116.         )
  118.         self.global_att = nn.Sequential(
  119.             nn.AdaptiveAvgPool2d(1),
  120.             nn.Conv2d(channels, inter_channels, kernel_size=1, stride=1, padding=0),
  121.             nn.BatchNorm2d(inter_channels),
  122.             nn.ReLU(inplace=True),
  123.             nn.Conv2d(inter_channels, channels, kernel_size=1, stride=1, padding=0),
  124.             nn.BatchNorm2d(channels),
  125.         )
  127.         self.sigmoid = nn.Sigmoid()
  129.     def forward(self, x, residual):
  130.         xa = x + residual
  131.         xl = self.local_att(xa)
  132.         xg = self.global_att(xa)
  133.         xlg = xl + xg
  134.         wei = self.sigmoid(xlg)
  136.         xo = 2 * x * wei + 2 * residual * (1 - wei)
  137.         return xo
  139. class Bottleneck_iAFF(nn.Module):
  140.     """Standard bottleneck."""
  142.     def __init__(self, c1, c2, shortcut=True, g=1, k=(3, 3), e=0.5):
  143.         """Initializes a bottleneck module with given input/output channels, shortcut option, group, kernels, and
  144.         expansion.
  145.         """
  146.         super().__init__()
  147.         c_ = int(c2 * e)  # hidden channels
  148.         self.cv1 = Conv(c1, c_, k[0], 1)
  149.         self.cv2 = Conv(c_, c2, k[1], 1, g=g)
  150.         self.add = shortcut and c1 == c2
  151.         self.iAFF = iAFF(c2)
  152.     def forward(self, x):
  153.         """'forward()' applies the YOLO FPN to input data."""
  154.         if self.add:
  155.            results = self.iAFF(x , self.cv2(self.cv1(x)))
  156.         else:
  157.             results = self.cv2(self.cv1(x))
  158.         return results
  160. class Bottleneck_AFF(nn.Module):
  161.     """Standard bottleneck."""
  163.     def __init__(self, c1, c2, shortcut=True, g=1, k=(3, 3), e=0.5):
  164.         """Initializes a bottleneck module with given input/output channels, shortcut option, group, kernels, and
  165.         expansion.
  166.         """
  167.         super().__init__()
  168.         c_ = int(c2 * e)  # hidden channels
  169.         self.cv1 = Conv(c1, c_, k[0], 1)
  170.         self.cv2 = Conv(c_, c2, k[1], 1, g=g)
  171.         self.add = shortcut and c1 == c2
  172.         self.AFF = AFF(c2)
  174.     def forward(self, x):
  175.         """'forward()' applies the YOLO FPN to input data."""
  176.         if self.add:
  177.             results = self.AFF(x, self.cv2(self.cv1(x)))
  178.         else:
  179.             results = self.cv2(self.cv1(x))
  180.         return results
  183. class Bottleneck(nn.Module):
  184.     """Standard bottleneck."""
  186.     def __init__(self, c1, c2, shortcut=True, g=1, k=(3, 3), e=0.5):
  187.         """Initializes a standard bottleneck module with optional shortcut connection and configurable parameters."""
  188.         super().__init__()
  189.         c_ = int(c2 * e)  # hidden channels
  190.         self.cv1 = Conv(c1, c_, k[0], 1)
  191.         self.cv2 = Conv(c_, c2, k[1], 1, g=g)
  192.         self.add = shortcut and c1 == c2
  194.     def forward(self, x):
  195.         """Applies the YOLO FPN to input data."""
  196.         return x + self.cv2(self.cv1(x)) if self.add else self.cv2(self.cv1(x))
  199. class C2f(nn.Module):
  200.     """Faster Implementation of CSP Bottleneck with 2 convolutions."""
  202.     def __init__(self, c1, c2, n=1, shortcut=False, g=1, e=0.5):
  203.         """Initializes a CSP bottleneck with 2 convolutions and n Bottleneck blocks for faster processing."""
  204.         super().__init__()
  205.         self.c = int(c2 * e)  # hidden channels
  206.         self.cv1 = Conv(c1, 2 * self.c, 1, 1)
  207.         self.cv2 = Conv((2 + n) * self.c, c2, 1)  # optional act=FReLU(c2)
  208.         self.m = nn.ModuleList(Bottleneck(self.c, self.c, shortcut, g, k=((3, 3), (3, 3)), e=1.0) for _ in range(n))
  210.     def forward(self, x):
  211.         """Forward pass through C2f layer."""
  212.         y = list(self.cv1(x).chunk(2, 1))
  213.         y.extend(m(y[-1]) for m in self.m)
  214.         return self.cv2(torch.cat(y, 1))
  216.     def forward_split(self, x):
  217.         """Forward pass using split() instead of chunk()."""
  218.         y = list(self.cv1(x).split((self.c, self.c), 1))
  219.         y.extend(m(y[-1]) for m in self.m)
  220.         return self.cv2(torch.cat(y, 1))
  223. class C3(nn.Module):
  224.     """CSP Bottleneck with 3 convolutions."""
  226.     def __init__(self, c1, c2, n=1, shortcut=True, g=1, e=0.5):
  227.         """Initialize the CSP Bottleneck with given channels, number, shortcut, groups, and expansion values."""
  228.         super().__init__()
  229.         c_ = int(c2 * e)  # hidden channels
  230.         self.cv1 = Conv(c1, c_, 1, 1)
  231.         self.cv2 = Conv(c1, c_, 1, 1)
  232.         self.cv3 = Conv(2 * c_, c2, 1)  # optional act=FReLU(c2)
  233.         self.m = nn.Sequential(*(Bottleneck(c_, c_, shortcut, g, k=((1, 1), (3, 3)), e=1.0) for _ in range(n)))
  235.     def forward(self, x):
  236.         """Forward pass through the CSP bottleneck with 2 convolutions."""
  237.         return self.cv3(torch.cat((self.m(self.cv1(x)), self.cv2(x)), 1))
  240. class C3k(C3):
  241.     """C3k is a CSP bottleneck module with customizable kernel sizes for feature extraction in neural networks."""
  243.     def __init__(self, c1, c2, n=1, shortcut=True, g=1, e=0.5, k=3):
  244.         """Initializes the C3k module with specified channels, number of layers, and configurations."""
  245.         super().__init__(c1, c2, n, shortcut, g, e)
  246.         c_ = int(c2 * e)  # hidden channels
  247.         # self.m = nn.Sequential(*(RepBottleneck(c_, c_, shortcut, g, k=(k, k), e=1.0) for _ in range(n)))
  248.         self.m = nn.Sequential(*(Bottleneck(c_, c_, shortcut, g, k=(k, k), e=1.0) for _ in range(n)))
  250. class C3k2_iAFF(C2f):
  251.     """Faster Implementation of CSP Bottleneck with 2 convolutions."""
  253.     def __init__(self, c1, c2, n=1, c3k=False, e=0.5, g=1, shortcut=True):
  254.         """Initializes the C3k2 module, a faster CSP Bottleneck with 2 convolutions and optional C3k blocks."""
  255.         super().__init__(c1, c2, n, shortcut, g, e)
  256.         self.m = nn.ModuleList(
  257.             C3k(self.c, self.c, 2, shortcut, g) if c3k else Bottleneck_iAFF(self.c, self.c, shortcut, g)for _ in range(n)
  258.         )
  259.         # 解析利用Bottleneck_iAFF替换Bottneck

四、手把手教你添加C3k2 iAFF

4.1 修改一

第一还是建立文件，我们找到如下 ultralytics /nn文件夹下建立一个目录名字呢就是'Addmodules'文件夹( 用群内的文件的话已经有了无需新建) ！然后在其内部建立一个新的py文件将核心代码复制粘贴进去即可。

​

​

4.2 修改二

第二步我们在该目录下创建一个新的py文件名字为'__init__.py'( 用群内的文件的话已经有了无需新建) ，然后在其内部导入我们的检测头如下图所示。

​​

4.3 修改三

第三步我门中到如下文件'ultralytics/nn/tasks.py'进行导入和注册我们的模块( 用群内的文件的话已经有了无需重新导入直接开始第四步即可) ！

​​

4.4 修改四

按照我的添加在parse_model里添加即可。

​​

到此就修改完成了，大家可以复制下面的yaml文件运行。

五、正式训练

5.1 yaml文件1

训练信息：YOLO11-C3k2-iAFF summary: 450 layers, 2,636,114 parameters, 2,636,098 gradients, 6.6 GFLOPs 

  1. # Ultralytics YOLO 🚀, AGPL-3.0 license
  2. # YOLO11 object detection model with P3-P5 outputs. For Usage examples see https://docs.ultralytics.com/tasks/detect
  4. # Parameters
  5. nc: 80 # number of classes
  6. scales: # model compound scaling constants, i.e. 'model=yolo11n.yaml' will call yolo11.yaml with scale 'n'
  7.   # [depth, width, max_channels]
  8.   n: [0.50, 0.25, 1024] # summary: 319 layers, 2624080 parameters, 2624064 gradients, 6.6 GFLOPs
  9.   s: [0.50, 0.50, 1024] # summary: 319 layers, 9458752 parameters, 9458736 gradients, 21.7 GFLOPs
  10.   m: [0.50, 1.00, 512] # summary: 409 layers, 20114688 parameters, 20114672 gradients, 68.5 GFLOPs
  11.   l: [1.00, 1.00, 512] # summary: 631 layers, 25372160 parameters, 25372144 gradients, 87.6 GFLOPs
  12.   x: [1.00, 1.50, 512] # summary: 631 layers, 56966176 parameters, 56966160 gradients, 196.0 GFLOPs
  14. # YOLO11n backbone
  15. backbone:
  16.   # [from, repeats, module, args]
  17.   - [-1, 1, Conv, [64, 3, 2]] # 0-P1/2
  18.   - [-1, 1, Conv, [128, 3, 2]] # 1-P2/4
  19.   - [-1, 2, C3k2_iAFF, [256, False, 0.25]]
  20.   - [-1, 1, Conv, [256, 3, 2]] # 3-P3/8
  21.   - [-1, 2, C3k2_iAFF, [512, False, 0.25]]
  22.   - [-1, 1, Conv, [512, 3, 2]] # 5-P4/16
  23.   - [-1, 2, C3k2_iAFF, [512, True]]
  24.   - [-1, 1, Conv, [1024, 3, 2]] # 7-P5/32
  25.   - [-1, 2, C3k2_iAFF, [1024, True]]
  26.   - [-1, 1, SPPF, [1024, 5]] # 9
  27.   - [-1, 2, C2PSA, [1024]] # 10
  29. # YOLO11n head
  30. head:
  31.   - [-1, 1, nn.Upsample, [None, 2, "nearest"]]
  32.   - [[-1, 6], 1, Concat, [1]] # cat backbone P4
  33.   - [-1, 2, C3k2_iAFF, [512, False]] # 13
  35.   - [-1, 1, nn.Upsample, [None, 2, "nearest"]]
  36.   - [[-1, 4], 1, Concat, [1]] # cat backbone P3
  37.   - [-1, 2, C3k2_iAFF, [256, False]] # 16 (P3/8-small)
  39.   - [-1, 1, Conv, [256, 3, 2]]
  40.   - [[-1, 13], 1, Concat, [1]] # cat head P4
  41.   - [-1, 2, C3k2_iAFF, [512, False]] # 19 (P4/16-medium)
  43.   - [-1, 1, Conv, [512, 3, 2]]
  44.   - [[-1, 10], 1, Concat, [1]] # cat head P5
  45.   - [-1, 2, C3k2_iAFF, [1024, True]] # 22 (P5/32-large)
  47.   - [[16, 19, 22], 1, Detect, [nc]] # Detect(P3, P4, P5)

5.2 训练代码

大家可以创建一个py文件将我给的代码复制粘贴进去，配置好自己的文件路径即可运行。 

  1. import warnings
  2. warnings.filterwarnings('ignore')
  3. from ultralytics import YOLO
  5. if __name__ == '__main__':
  6.     model = YOLO('模型配置文件')
  7.     # 如何切换模型版本, 上面的ymal文件可以改为 yolov8s.yaml就是使用的v8s,
  8.     # 类似某个改进的yaml文件名称为yolov8-XXX.yaml那么如果想使用其它版本就把上面的名称改为yolov8l-XXX.yaml即可（改的是上面YOLO中间的名字不是配置文件的）！
  9.     # model.load('yolov8n.pt') # 是否加载预训练权重,科研不建议大家加载否则很难提升精度
  10.     model.train(data=r"C:\Users\Administrator\PycharmProjects\yolov5-master\yolov5-master\Construction Site Safety.v30-raw-images_latestversion.yolov8\data.yaml",
  11.                 # 如果大家任务是其它的'ultralytics/cfg/default.yaml'找到这里修改task可以改成detect, segment, classify, pose
  12.                 cache=False,
  13.                 imgsz=640,
  14.                 epochs=150,
  15.                 single_cls=False,  # 是否是单类别检测
  16.                 batch=16,
  17.                 close_mosaic=0,
  18.                 workers=0,
  19.                 device='0',
  20.                 optimizer='SGD', # using SGD
  21.                 # resume='runs/train/exp21/weights/last.pt', # 如过想续训就设置last.pt的地址
  22.                 amp=True,  # 如果出现训练损失为Nan可以关闭amp
  23.                 project='runs/train',
  24.                 name='exp',
  25.                 )

5.3 训练过程截图

​

五、本文总结

到此本文的正式分享内容就结束了，在这里给大家推荐我的YOLOv11改进有效涨点专栏，本专栏目前为新开的平均质量分98分，后期我会根据各种最新的前沿顶会进行论文复现，也会对一些老的改进机制进行补充，如果大家觉得本文帮助到你了，订阅本专栏，关注后续更多的更新~