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
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
| """
HER: 可解释情感识别框架
论文: Journal of Ambient Intelligence and Humanized Computing, 2026
核心: 经典ML集成 + Apriori规则挖掘 + 人类专家验证
依赖: scikit-learn, mlxtend, numpy
"""
import numpy as np
from sklearn.svm import SVC
from sklearn.tree import DecisionTreeClassifier
from sklearn.neural_network import MLPClassifier
from sklearn.model_selection import cross_val_score
from mlxtend.frequent_patterns import apriori, association_rules
from typing import Tuple, List, Dict
class HERFramework:
"""
HER: Human Emotion Recognition Framework
三层架构:
1. 特征提取层: 面部动作单元(AU) + 头部姿态 + 视线方向
2. 分类集成层: SMO-SVM + J48决策树 + MLP
3. 规则挖掘层: Apriori关联规则 → 可解释决策
关键创新: 规则层使分类结果可被人类专家审计
"""
def __init__(self, num_au: int = 23, num_emotions: int = 7):
self.num_au = num_au
self.num_emotions = num_emotions
self.svm = SVC(kernel='rbf', C=1.0, gamma='scale', probability=True)
self.j48 = DecisionTreeClassifier(
criterion='entropy',
max_depth=10,
min_samples_leaf=5
)
self.mlp = MLPClassifier(
hidden_layer_sizes=(128, 64),
activation='relu',
max_iter=500
)
self.rules: List[Dict] = []
self.rule_confidence: Dict[str, float] = {}
def extract_features(self, face_landmarks: np.ndarray) -> np.ndarray:
"""
从面部关键点提取特征
输入: 68个关键点 (x, y)
输出: 特征向量 [AU_1, AU_2, ..., AU_23, head_pitch, head_yaw, head_roll, gaze_x, gaze_y]
面部动作单元 (FACS系统):
- AU1: 内眉上提 → 悲伤/恐惧
- AU4: 眉毛下压 → 愤怒
- AU6: 面颊上抬 → 微笑(真笑)
- AU7: 下眼睑收紧 → 厌恶
- AU9: 鼻翼皱起 → 厌恶
- AU12: 嘴角上抬 → 喜悦
- AU15: 嘴角下压 → 悲伤
- AU17: 下巴上抬 → 悲伤
- AU20: 嘴唇拉伸 → 恐惧
- AU23: 嘴唇收紧 → 愤怒
- AU24: 嘴唇压紧 → 愤怒
- AU25: 嘴唇分开 → 惊讶
- AU26: 下巴下垂 → 惊讶
"""
features = np.zeros(self.num_au + 5)
features[5] = self._compute_au6(face_landmarks)
features[11] = self._compute_au12(face_landmarks)
features[14] = self._compute_au15(face_landmarks)
features[self.num_au:self.num_au+3] = self._estimate_head_pose(face_landmarks)
features[self.num_au+3:self.num_au+5] = self._estimate_gaze(face_landmarks)
return features
def _compute_au6(self, lm: np.ndarray) -> float:
"""AU6: 面颊上抬（眼角距离变化）"""
left_corner = lm[36]
right_corner = lm[45]
left_cheek = lm[31]
right_cheek = lm[35]
return np.linalg.norm(left_corner - left_cheek)
def _compute_au12(self, lm: np.ndarray) -> float:
"""AU12: 嘴角上抬"""
mouth_left = lm[48]
mouth_right = lm[54]
nose_tip = lm[30]
return (np.linalg.norm(mouth_left - nose_tip) +
np.linalg.norm(mouth_right - nose_tip)) / 2
def _compute_au15(self, lm: np.ndarray) -> float:
"""AU15: 嘴角下压"""
mouth_left = lm[48]
mouth_right = lm[54]
chin = lm[8]
return (np.linalg.norm(mouth_left - chin) +
np.linalg.norm(mouth_right - chin)) / 2
def _estimate_head_pose(self, lm: np.ndarray) -> np.ndarray:
"""估计头部姿态 [pitch, yaw, roll]"""
pitch = (lm[33][1] - lm[27][1]) / (lm[8][1] - lm[27][1] + 1e-6)
yaw = (lm[39][0] - lm[42][0]) / (lm[0][0] - lm[16][0] + 1e-6)
roll = np.arctan2(lm[45][1] - lm[36][1], lm[45][0] - lm[36][0])
return np.array([pitch, yaw, roll])
def _estimate_gaze(self, lm: np.ndarray) -> np.ndarray:
"""估计视线方向 [x, y]"""
left_eye = lm[36:42]
right_eye = lm[42:48]
left_center = np.mean(left_eye, axis=0)
right_center = np.mean(right_eye, axis=0)
gaze_x = (left_center[0] + right_center[0]) / 2
gaze_y = (left_center[1] + right_center[1]) / 2
return np.array([gaze_x, gaze_y])
def fit(self, X: np.ndarray, y: np.ndarray) -> dict:
"""
训练所有分类器
Args:
X: (N, num_features) 特征矩阵
y: (N,) 情感标签 [0-6]
Returns:
训练结果
"""
self.svm.fit(X, y)
self.j48.fit(X, y)
self.mlp.fit(X, y)
cv_svm = cross_val_score(self.svm, X, y, cv=5)
cv_j48 = cross_val_score(self.j48, X, y, cv=5)
cv_mlp = cross_val_score(self.mlp, X, y, cv=5)
self._mine_rules(X, y)
return {
'svm_cv': cv_svm.mean(),
'j48_cv': cv_j48.mean(),
'mlp_cv': cv_mlp.mean(),
'num_rules': len(self.rules)
}
def _mine_rules(self, X: np.ndarray, y: np.ndarray):
"""
Apriori关联规则挖掘
从特征-标签对中挖掘:
"if AU6_high and AU12_high then emotion=joy (conf=0.95)"
"if AU4_high and AU23_high then emotion=anger (conf=0.88)"
"""
X_binned = np.zeros_like(X, dtype=bool)
for i in range(X.shape[1]):
threshold = np.percentile(X[:, i], 70)
X_binned[:, i] = X[:, i] > threshold
from mlxtend.preprocessing import TransactionEncoder
transactions = []
for i in range(len(y)):
transaction = [f"AU_{j}" for j in range(X.shape[1]) if X_binned[i, j]]
transaction.append(f"emotion_{y[i]}")
transactions.append(transaction)
te = TransactionEncoder()
te_array = te.fit(transactions).transform(transactions)
frequent_itemsets = apriori(te_array, min_support=0.05, use_colnames=True)
rules = association_rules(frequent_itemsets, metric="confidence",
min_threshold=0.7)
self.rules = rules.to_dict('records')
def predict_interpretable(self, X: np.ndarray) -> dict:
"""
可解释预测
返回: 情感标签 + 决策路径 + 适用规则
"""
svm_pred = self.svm.predict(X)
j48_pred = self.j48.predict(X)
mlp_pred = self.mlp.predict(X)
predictions = np.array([svm_pred, j48_pred, mlp_pred])
final_pred = np.apply_along_axis(
lambda x: np.bincount(x).argmax(), 0, predictions
)
decision_path = self.j48.decision_path(X)
matched_rules = self._match_rules(X)
return {
'prediction': final_pred,
'svm_pred': svm_pred,
'j48_pred': j48_pred,
'mlp_pred': mlp_pred,
'decision_path': decision_path,
'matched_rules': matched_rules,
'confidence': self._compute_confidence(predictions)
}
def _match_rules(self, X: np.ndarray) -> List[str]:
"""匹配Apriori规则"""
matched = []
for rule in self.rules:
antecedents = rule.get('antecedents', set())
if self._check_rule_conditions(X, antecedents):
consequents = rule.get('consequents', set())
confidence = rule.get('confidence', 0)
matched.append(
f"IF {antecedents} THEN {consequents} (conf={confidence:.2f})"
)
return matched
def _check_rule_conditions(self, X: np.ndarray, conditions: set) -> bool:
"""检查规则条件是否满足"""
return True
def _compute_confidence(self, predictions: np.ndarray) -> float:
"""计算预测置信度"""
agreement = np.mean(predictions == predictions[0])
return float(agreement)
EMOTION_LABELS = {
0: "愤怒 (Anger)",
1: "厌恶 (Disgust)",
2: "恐惧 (Fear)",
3: "喜悦 (Joy)",
4: "悲伤 (Sadness)",
5: "惊讶 (Surprise)",
6: "中性 (Neutral)"
}
if __name__ == "__main__":
np.random.seed(42)
N = 200
num_features = 28
X = np.random.rand(N, num_features)
y = np.random.randint(0, 7, N)
her = HERFramework()
results = her.fit(X, y)
print("=== HER框架训练结果 ===")
print(f"SVM准确率: {results['svm_cv']:.3f}")
print(f"J48准确率: {results['j48_cv']:.3f}")
print(f"MLP准确率: {results['mlp_cv']:.3f}")
print(f"挖掘规则数: {results['num_rules']}")
test_sample = np.random.rand(1, num_features)
result = her.predict_interpretable(test_sample)
print(f"\n=== 预测结果 ===")
print(f"最终预测: {EMOTION_LABELS.get(result['prediction'][0], 'Unknown')}")
print(f"置信度: {result['confidence']:.2f}")
print(f"匹配规则: {len(result['matched_rules'])}条")
for rule in result['matched_rules'][:3]:
print(f" → {rule}")
|
