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
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
| """
Guided Dual-Attention Network: 13种注意力机制对比
论文: Technologies (MDPI), 2026, 14(9), 544
核心: 双注意力机制
- Spatial Attention (空间注意力): 聚焦眼睛关键区域
- Channel Attention (通道注意力): 聚焦重要特征通道
- Guided: 空间和通道注意力互相引导
13种变体:
1. SE (Squeeze-and-Excitation)
2. CBAM (Convolutional Block Attention Module)
3. ECA (Efficient Channel Attention)
4. CA (Coordinate Attention)
5. GA (Gated Attention)
6. TA (Triplet Attention)
7. PA (Polarized Attention)
8. SA (Spatial Attention)
9. CA-only (Channel Only)
10. SA-only (Spatial Only)
11. SGE (Spatial Group-wise Enhancement)
12. CGA (Convolutional Guided Attention) ← 论文提出
13. Baseline (无注意力)
"""
import torch
import torch.nn as nn
import torch.nn.functional as F
class SpatialAttention(nn.Module):
"""空间注意力: 聚焦眼睛关键区域"""
def __init__(self, kernel_size=7):
super().__init__()
self.conv = nn.Conv2d(2, 1, kernel_size, padding=kernel_size//2)
def forward(self, x):
avg_out = torch.mean(x, dim=1, keepdim=True)
max_out, _ = torch.max(x, dim=1, keepdim=True)
cat = torch.cat([avg_out, max_out], dim=1)
return x * torch.sigmoid(self.conv(cat))
class ChannelAttention(nn.Module):
"""通道注意力: 聚焦重要特征通道"""
def __init__(self, channels, reduction=16):
super().__init__()
self.avg_pool = nn.AdaptiveAvgPool2d(1)
self.max_pool = nn.AdaptiveMaxPool2d(1)
self.mlp = nn.Sequential(
nn.Linear(channels, channels // reduction),
nn.ReLU(),
nn.Linear(channels // reduction, channels)
)
def forward(self, x):
b, c, _, _ = x.shape
avg_out = self.mlp(self.avg_pool(x).view(b, c))
max_out = self.mlp(self.max_pool(x).view(b, c))
return x * torch.sigmoid(avg_out + max_out).view(b, c, 1, 1)
class GuidedDualAttention(nn.Module):
"""
论文提出的CGA (Convolutional Guided Attention)
核心: 空间和通道注意力互相引导
- 通道注意力输出 → 引导空间注意力
- 空间注意力输出 → 引导通道注意力
- 双向引导形成闭环
"""
def __init__(self, channels, reduction=16, kernel_size=7):
super().__init__()
self.ca = ChannelAttention(channels, reduction)
self.sa = SpatialAttention(kernel_size)
self.gate_c = nn.Sequential(
nn.Conv2d(channels, 1, 1),
nn.Sigmoid()
)
self.gate_s = nn.Sequential(
nn.Conv2d(1, channels, 1),
nn.Sigmoid()
)
def forward(self, x):
ca_out = self.ca(x)
sa_out = self.sa(ca_out)
c_guide = self.gate_c(ca_out)
s_guide = self.gate_s(sa_out.mean(dim=1, keepdim=True))
out = x * ca_out * s_guide + x * sa_out * c_guide
return out
class DrowsinessDetector(nn.Module):
"""
完整疲劳检测模型
输入: 眼部ROI图像 (48x48x3)
输出: 疲劳概率 (0-1)
部署: FPGA INT8量化
"""
def __init__(self, attention_type='cga'):
super().__init__()
self.features = nn.Sequential(
nn.Conv2d(3, 32, 3, padding=1),
nn.BatchNorm2d(32),
nn.ReLU(),
nn.MaxPool2d(2),
nn.Conv2d(32, 64, 3, padding=1),
nn.BatchNorm2d(64),
nn.ReLU(),
nn.MaxPool2d(2),
)
if attention_type == 'cga':
self.attention = GuidedDualAttention(64)
elif attention_type == 'se':
self.attention = ChannelAttention(64)
elif attention_type == 'sa':
self.attention = SpatialAttention()
else:
self.attention = nn.Identity()
self.classifier = nn.Sequential(
nn.AdaptiveAvgPool2d(1),
nn.Flatten(),
nn.Dropout(0.3),
nn.Linear(64, 32),
nn.ReLU(),
nn.Linear(32, 2)
)
def forward(self, x):
x = self.features(x)
x = self.attention(x)
return self.classifier(x)
if __name__ == "__main__":
model = DrowsinessDetector('cga')
x = torch.randn(1, 3, 48, 48)
out = model(x)
params = sum(p.numel() for p in model.parameters())
print(f"模型输出: {out.shape}")
print(f"参数量: {params:,} ({params/1e3:.1f}K)")
print(f"FP32大小: {params*4/1024:.1f}KB")
print(f"INT8大小: {params/1024:.1f}KB")
|
