一、本文介绍
本文记录的是基于NAM模块的YOLOv11目标检测改进方法研究。 许多先前的研究专注于通过注意力操作捕获显著特征,但缺乏对权重贡献因素的考虑,而这些因素能够进一步抑制不重要的通道或像素。而本文利用NAM
改进YOLOv11
,通过权重的贡献因素来改进注意力机制,提高模型精度。
文章目录
二、NAM介绍
NAM: Normalization-based Attention Module
NAM(Normalization - based Attention Module)
注意力模块的设计的原理和优势如下:
2.1 NAM设计原理
NAM
采用了来自CBAM(Convolutional Block Attention Module)
的模块集成方式,并重新设计了通道和空间注意力子模块。- 在通道注意力子模块中,使用了批归一化(Batch Normalization,BN)的缩放因子来衡量通道的方差,并表示其重要性。具体公式为: B o u t = B N ( B i n ) = γ B i n − μ B σ B 2 + ϵ + β B_{out } = BN(B_{in}) = \gamma \frac{B_{in} - \mu_{\mathcal{B}}}{\sqrt{\sigma_{\mathcal{B}}^{2} + \epsilon}} + \beta Bout=BN(Bin)=γσB2+ϵBin−μB+β,其中 μ B \mu_{B} μB和 σ B \sigma_{B} σB分别是小批量 B B B的均值和标准差; γ \gamma γ和 β \beta β是可训练的仿射变换参数(缩放和平移)。通道注意力子模块的输出特征 M c M_{c} Mc表示为: M c = s i g m o i d ( W γ ( B N ( F 1 ) ) ) M_{c} = sigmoid(W_{\gamma}(BN(F_{1}))) Mc=sigmoid(Wγ(BN(F1))),其中 γ \gamma γ是每个通道的缩放因子,权重 W γ W_{\gamma} Wγ通过 W γ = γ i / ∑ j = 0 γ j W_{\gamma} = \gamma_{i} / \sum_{j = 0} \gamma_{j} Wγ=γi/∑j=0γj获得。
- 在空间维度上也应用了BN的缩放因子来测量像素的重要性,称为像素归一化。相应的空间注意力子模块的输出 M s M_{s} Ms表示为: M s = s i g m o i d ( W λ ( B N s ( F 2 ) ) ) M_{s} = sigmoid(W_{\lambda}(BN_{s}(F_{2}))) Ms=sigmoid(Wλ(BNs(F2))),其中 X X X是缩放因子,权重 W λ W_{\lambda} Wλ通过 W λ = λ i / ∑ j = 0 λ j W_{\lambda} = \lambda_{i} / \sum_{j = 0} \lambda_{j} Wλ=λi/∑j=0λj获得。
- 为了抑制不太显著的权重,在损失函数中添加了一个正则化项,具体公式为: L o s s = ∑ ( x , y ) l ( f ( x , W ) , y ) + p ∑ g ( γ ) + p ∑ g ( λ ) Loss = \sum_{(x, y)} l(f(x, W), y) + p \sum g(\gamma) + p \sum g(\lambda) Loss=∑(x,y)l(f(x,W),y)+p∑g(γ)+p∑g(λ),其中 x x x表示输入, y y y是输出, W W W代表网络权重, l ( ⋅ ) l(\cdot) l(⋅)是损失函数, g ( − ) g(-) g(−)是 l 1 l_{1} l1范数惩罚函数, p p p是平衡 g ( γ ) g(\gamma) g(γ)和 g ( λ ) g(\lambda) g(λ)的惩罚项。
2.2 优势
- 通过抑制不太显著的特征,
NAM
更高效。 - 与其他三种注意力机制(SE、BAM、CBAM)在ResNet和MobileNet上的比较表明,
NAM
在单独使用通道或空间注意力时,性能优于其他四种注意力机制;在结合通道和空间注意力时,在具有相似计算复杂度的情况下,性能也更优。 - 与CBAM相比,
NAM
在通道注意力模块中显著减少了参数数量,在空间注意力模块中参数增加不显著,总体上参数更少。
论文:https://arxiv.org/pdf/2111.12419
源码:https://github.com/Christian-lyc/NAM
三、NAM的实现代码
NAM模块
的实现代码如下:
class Channel_Att(nn.Module):
def __init__(self, channels, t=16):
super(Channel_Att, self).__init__()
self.channels = channels
self.bn2 = nn.BatchNorm2d(self.channels, affine=True)
def forward(self, x):
residual = x
x = self.bn2(x)
weight_bn = self.bn2.weight.data.abs() / torch.sum(self.bn2.weight.data.abs())
x = x.permute(0, 2, 3, 1).contiguous()
x = torch.mul(weight_bn, x)
x = x.permute(0, 3, 1, 2).contiguous()
x = torch.sigmoid(x) * residual #
return x
class NAM(nn.Module):
def __init__(self, channels, out_channels=None, no_spatial=True):
super(NAM, self).__init__()
self.Channel_Att = Channel_Att(channels)
def forward(self, x):
x_out1=self.Channel_Att(x)
return x_out1
四、创新模块
4.1 改进点1
模块改进方法1️⃣:直接加入NAM模块
。
NAM模块
添加后如下:
注意❗:在5.2和5.3小节
中需要声明的模块名称为:NAM
。
4.2 二次创新⭐
模块改进方法2️⃣:基于NAM模块
的C3k2
。
📌 第二种改进方法是对YOLOv11
中的C3k2模块
进行改进,其中在C2f
提取特征后,利用NAM
重新设计通道和空间注意力子模块,从而抑制不太显著的特征,并且在与C3k2
结合后,对于细节特征的提取更加敏感,提高模型性能。
改进代码如下:
首先添加如下代码改进C2f模块
,并重命名为C2f_NAM
class C2f_NAM(nn.Module):
"""Faster Implementation of CSP Bottleneck with 2 convolutions."""
def __init__(self, c1, c2, n=1, shortcut=False, g=1, e=0.5):
"""Initialize CSP bottleneck layer with two convolutions with arguments ch_in, ch_out, number, shortcut, groups,
expansion.
"""
super().__init__()
self.c = int(c2 * e) # hidden channels
self.cv1 = Conv(c1, 2 * self.c, 1, 1)
self.cv2 = Conv((2 + n) * self.c, c2, 1) # optional act=FReLU(c2)
self.m = nn.ModuleList(Bottleneck(self.c, self.c, shortcut, g, k=((3, 3), (3, 3)), e=1.0) for _ in range(n))
self.att = NAM(c2)
def forward(self, x):
"""Forward pass through C2f layer."""
y = list(self.cv1(x).chunk(2, 1))
y.extend(m(y[-1]) for m in self.m)
return self.att(self.cv2(torch.cat(y, 1)))
def forward_split(self, x):
"""Forward pass using split() instead of chunk()."""
y = list(self.cv1(x).split((self.c, self.c), 1))
y.extend(m(y[-1]) for m in self.m)
return self.att(self.cv2(torch.cat(y, 1)))
再添加如下代码让C3k2
继承于C2f_NAM
,并重命名为C3k2_NAM
class C3k2_NAM(C2f_NAM):
"""Faster Implementation of CSP Bottleneck with 2 convolutions."""
def __init__(self, c1, c2, n=1, c3k=False, e=0.5, g=1, shortcut=True):
"""Initializes the C3k2 module, a faster CSP Bottleneck with 2 convolutions and optional C3k blocks."""
super().__init__(c1, c2, n, shortcut, g, e)
self.m = nn.ModuleList(
C3k(self.c, self.c, 2, shortcut, g) if c3k else Bottleneck(self.c, self.c, shortcut, g) for _ in range(n)
)
注意❗:在5.2和5.3小节
中需要声明的模块名称为:C3k2_NAM
。
五、添加步骤
5.1 修改ultralytics/nn/modules/block.py
此处需要修改的文件是ultralytics/nn/modules/block.py
block.py中定义了网络结构的通用模块
,我们想要加入新的模块就只需要将模块代码放到这个文件内即可。
将NAM
和C3k2_NAM
模块代码添加到此文件下。
5.2 修改ultralytics/nn/modules/init.py
此处需要修改的文件是ultralytics/nn/modules/__init__.py
__init__.py
文件中定义了所有模块的初始化,我们只需要将block.py
中的新的模块命添加到对应的函数即可。
NAM
和C2f_NAM
在block.py
中实现,所有要添加在from .block import
:
from .block import (
C1,
C2,
...
NAM,
C3k2_NAM
)
5.3 修改ultralytics/nn/modules/tasks.py
在tasks.py
文件中,需要在两处位置添加各模块类名称。
首先:在函数声明中引入NAM
和C3k2_NAM
其次:在parse_model函数
中注册NAM
和C3k2_NAM
模块
六、yaml模型文件
6.1 模型改进版本一
在代码配置完成后,配置模型的YAML文件。
此处以ultralytics/cfg/models/11/yolov11m.yaml
为例,在同目录下创建一个用于自己数据集训练的模型文件yolov11m-NAM.yaml
。
将yolov11m.yaml
中的内容复制到yolov11m-NAM.yaml
文件下,修改nc
数量等于自己数据中目标的数量。
在骨干网络的最后一层添加NAM模块
,只需要填入一个参数,通道数,和前一层通道数一致。还需要注意的是,由于PAN+FPN的颈部模型结构存在,层之间的匹配也要记得修改,维度要匹配上。
# Ultralytics YOLO 🚀, AGPL-3.0 license
# YOLO11 object detection model with P3-P5 outputs. For Usage examples see https://docs.ultralytics.com/tasks/detect
# yolo task=detect mode=train model=yolov11m.yaml data=data.yaml device=0 epochs=300 batch=16 imgsz=640 workers=10
# Parameters
nc: 1 # number of classes
scales: # model compound scaling constants, i.e. 'model=yolo11n.yaml' will call yolo11.yaml with scale 'n'
# [depth, width, max_channels]
n: [0.50, 0.25, 1024] # summary: 319 layers, 2624080 parameters, 2624064 gradients, 6.6 GFLOPs
s: [0.50, 0.50, 1024] # summary: 319 layers, 9458752 parameters, 9458736 gradients, 21.7 GFLOPs
m: [0.50, 1.00, 512] # summary: 409 layers, 20114688 parameters, 20114672 gradients, 68.5 GFLOPs
l: [1.00, 1.00, 512] # summary: 631 layers, 25372160 parameters, 25372144 gradients, 87.6 GFLOPs
x: [1.00, 1.50, 512] # summary: 631 layers, 56966176 parameters, 56966160 gradients, 196.0 GFLOPs
# YOLO11n backbone
backbone:
# [from, repeats, module, args]
- [-1, 1, Conv, [64, 3, 2]] # 0-P1/2
- [-1, 1, Conv, [128, 3, 2]] # 1-P2/4
- [-1, 2, C3k2, [256, False, 0.25]]
- [-1, 1, Conv, [256, 3, 2]] # 3-P3/8
- [-1, 2, C3k2, [512, False, 0.25]]
- [-1, 1, Conv, [512, 3, 2]] # 5-P4/16
- [-1, 2, C3k2, [512, True]]
- [-1, 1, Conv, [1024, 3, 2]] # 7-P5/32
- [-1, 2, C3k2, [1024, True]]
- [-1, 1, NAM, [1024]]
- [-1, 1, SPPF, [1024, 5]] # 9
- [-1, 2, C2PSA, [1024]] # 10
# YOLO11n head
head:
- [-1, 1, nn.Upsample, [None, 2, "nearest"]]
- [[-1, 6], 1, Concat, [1]] # cat backbone P4
- [-1, 2, C3k2, [512, False]] # 13
- [-1, 1, nn.Upsample, [None, 2, "nearest"]]
- [[-1, 4], 1, Concat, [1]] # cat backbone P3
- [-1, 2, C3k2, [256, False]] # 16 (P3/8-small)
- [-1, 1, Conv, [256, 3, 2]]
- [[-1, 14], 1, Concat, [1]] # cat head P4
- [-1, 2, C3k2, [512, False]] # 19 (P4/16-medium)
- [-1, 1, Conv, [512, 3, 2]]
- [[-1, 11], 1, Concat, [1]] # cat head P5
- [-1, 2, C3k2, [1024, True]] # 22 (P5/32-large)
- [[17, 20, 23], 1, Detect, [nc]] # Detect(P3, P4, P5)
6.2 模型改进版本二⭐
此处同样以ultralytics/cfg/models/11/yolov11m.yaml
为例,在同目录下创建一个用于自己数据集训练的模型文件yolov11m-C3k2_NAM.yaml
。
将yolov11m.yaml
中的内容复制到yolov11m-C3k2_NAM.yaml
文件下,修改nc
数量等于自己数据中目标的数量。
📌 模型的修改方法是将骨干网络中的所有C3k2模块
替换成C3k2_NAM模块
。
# Ultralytics YOLO 🚀, AGPL-3.0 license
# YOLO11 object detection model with P3-P5 outputs. For Usage examples see https://docs.ultralytics.com/tasks/detect
# yolo task=detect mode=train model=yolov11m.yaml data=data.yaml device=0 epochs=300 batch=16 imgsz=640 workers=10
# Parameters
nc: 1 # number of classes
scales: # model compound scaling constants, i.e. 'model=yolo11n.yaml' will call yolo11.yaml with scale 'n'
# [depth, width, max_channels]
n: [0.50, 0.25, 1024] # summary: 319 layers, 2624080 parameters, 2624064 gradients, 6.6 GFLOPs
s: [0.50, 0.50, 1024] # summary: 319 layers, 9458752 parameters, 9458736 gradients, 21.7 GFLOPs
m: [0.50, 1.00, 512] # summary: 409 layers, 20114688 parameters, 20114672 gradients, 68.5 GFLOPs
l: [1.00, 1.00, 512] # summary: 631 layers, 25372160 parameters, 25372144 gradients, 87.6 GFLOPs
x: [1.00, 1.50, 512] # summary: 631 layers, 56966176 parameters, 56966160 gradients, 196.0 GFLOPs
# YOLO11n backbone
backbone:
# [from, repeats, module, args]
- [-1, 1, Conv, [64, 3, 2]] # 0-P1/2
- [-1, 1, Conv, [128, 3, 2]] # 1-P2/4
- [-1, 2, C3k2_NAM, [256, False, 0.25]]
- [-1, 1, Conv, [256, 3, 2]] # 3-P3/8
- [-1, 2, C3k2_NAM, [512, False, 0.25]]
- [-1, 1, Conv, [512, 3, 2]] # 5-P4/16
- [-1, 2, C3k2_NAM, [512, True]]
- [-1, 1, Conv, [1024, 3, 2]] # 7-P5/32
- [-1, 2, C3k2_NAM, [1024, True]]
- [-1, 1, SPPF, [1024, 5]] # 9
- [-1, 2, C2PSA, [1024]] # 10
# YOLO11n head
head:
- [-1, 1, nn.Upsample, [None, 2, "nearest"]]
- [[-1, 6], 1, Concat, [1]] # cat backbone P4
- [-1, 2, C3k2, [512, False]] # 13
- [-1, 1, nn.Upsample, [None, 2, "nearest"]]
- [[-1, 4], 1, Concat, [1]] # cat backbone P3
- [-1, 2, C3k2, [256, False]] # 16 (P3/8-small)
- [-1, 1, Conv, [256, 3, 2]]
- [[-1, 13], 1, Concat, [1]] # cat head P4
- [-1, 2, C3k2, [512, False]] # 19 (P4/16-medium)
- [-1, 1, Conv, [512, 3, 2]]
- [[-1, 10], 1, Concat, [1]] # cat head P5
- [-1, 2, C3k2, [1024, True]] # 22 (P5/32-large)
- [[16, 19, 22], 1, Detect, [nc]] # Detect(P3, P4, P5)
七、成功运行结果
分别打印网络模型可以看到NAM模块
和C3k2_NAM
已经加入到模型中,并可以进行训练了。
yolov11m-NAM:
from n params module arguments
0 -1 1 1856 ultralytics.nn.modules.conv.Conv [3, 64, 3, 2]
1 -1 1 73984 ultralytics.nn.modules.conv.Conv [64, 128, 3, 2]
2 -1 1 111872 ultralytics.nn.modules.block.C3k2 [128, 256, 1, True, 0.25]
3 -1 1 590336 ultralytics.nn.modules.conv.Conv [256, 256, 3, 2]
4 -1 1 444928 ultralytics.nn.modules.block.C3k2 [256, 512, 1, True, 0.25]
5 -1 1 2360320 ultralytics.nn.modules.conv.Conv [512, 512, 3, 2]
6 -1 1 1380352 ultralytics.nn.modules.block.C3k2 [512, 512, 1, True]
7 -1 1 2360320 ultralytics.nn.modules.conv.Conv [512, 512, 3, 2]
8 -1 1 1380352 ultralytics.nn.modules.block.C3k2 [512, 512, 1, True]
9 -1 1 1024 ultralytics.nn.modules.block.NAM [512, 512]
10 -1 1 656896 ultralytics.nn.modules.block.SPPF [512, 512, 5]
11 -1 1 990976 ultralytics.nn.modules.block.C2PSA [512, 512, 1]
12 -1 1 0 torch.nn.modules.upsampling.Upsample [None, 2, 'nearest']
13 [-1, 6] 1 0 ultralytics.nn.modules.conv.Concat [1]
14 -1 1 1642496 ultralytics.nn.modules.block.C3k2 [1024, 512, 1, True]
15 -1 1 0 torch.nn.modules.upsampling.Upsample [None, 2, 'nearest']
16 [-1, 4] 1 0 ultralytics.nn.modules.conv.Concat [1]
17 -1 1 542720 ultralytics.nn.modules.block.C3k2 [1024, 256, 1, True]
18 -1 1 590336 ultralytics.nn.modules.conv.Conv [256, 256, 3, 2]
19 [-1, 14] 1 0 ultralytics.nn.modules.conv.Concat [1]
20 -1 1 1511424 ultralytics.nn.modules.block.C3k2 [768, 512, 1, True]
21 -1 1 2360320 ultralytics.nn.modules.conv.Conv [512, 512, 3, 2]
22 [-1, 11] 1 0 ultralytics.nn.modules.conv.Concat [1]
23 -1 1 1642496 ultralytics.nn.modules.block.C3k2 [1024, 512, 1, True]
24 [17, 20, 23] 1 1411795 ultralytics.nn.modules.head.Detect [1, [256, 512, 512]]
YOLOv11m-NAM summary: 412 layers, 20,054,803 parameters, 20,054,787 gradients, 68.2 GFLOPs
yolov11m-C3k2_NAM:
from n params module arguments
0 -1 1 1856 ultralytics.nn.modules.conv.Conv [3, 64, 3, 2]
1 -1 1 73984 ultralytics.nn.modules.conv.Conv [64, 128, 3, 2]
2 -1 1 99840 ultralytics.nn.modules.block.C3k2_NAM [128, 256, False, 0.25]
3 -1 1 590336 ultralytics.nn.modules.conv.Conv [256, 256, 3, 2]
4 -1 1 396288 ultralytics.nn.modules.block.C3k2_NAM [256, 512, False, 0.25]
5 -1 1 2360320 ultralytics.nn.modules.conv.Conv [512, 512, 3, 2]
6 -1 1 1249024 ultralytics.nn.modules.block.C3k2_NAM [512, 512, True]
7 -1 1 2360320 ultralytics.nn.modules.conv.Conv [512, 512, 3, 2]
8 -1 1 1249024 ultralytics.nn.modules.block.C3k2_NAM [512, 512, True]
9 -1 1 656896 ultralytics.nn.modules.block.SPPF [512, 512, 5]
10 -1 1 990976 ultralytics.nn.modules.block.C2PSA [512, 512, 1]
11 -1 1 0 torch.nn.modules.upsampling.Upsample [None, 2, 'nearest']
12 [-1, 6] 1 0 ultralytics.nn.modules.conv.Concat [1]
13 -1 1 1642496 ultralytics.nn.modules.block.C3k2 [1024, 512, 1, True]
14 -1 1 0 torch.nn.modules.upsampling.Upsample [None, 2, 'nearest']
15 [-1, 4] 1 0 ultralytics.nn.modules.conv.Concat [1]
16 -1 1 542720 ultralytics.nn.modules.block.C3k2 [1024, 256, 1, True]
17 -1 1 590336 ultralytics.nn.modules.conv.Conv [256, 256, 3, 2]
18 [-1, 13] 1 0 ultralytics.nn.modules.conv.Concat [1]
19 -1 1 1511424 ultralytics.nn.modules.block.C3k2 [768, 512, 1, True]
20 -1 1 2360320 ultralytics.nn.modules.conv.Conv [512, 512, 3, 2]
21 [-1, 10] 1 0 ultralytics.nn.modules.conv.Concat [1]
22 -1 1 1642496 ultralytics.nn.modules.block.C3k2 [1024, 512, 1, True]
23 [16, 19, 22] 1 1411795 ultralytics.nn.modules.head.Detect [1, [256, 512, 512]]
YOLOv11m-C3k2_NAM summary: 335 layers, 19,730,451 parameters, 19,730,435 gradients, 66.4 GFLOPs