Appearance
第二十四章 · DNN 模块深入
24.1 DNN 推理流程详解
24.1.1 标准推理管线
python
import cv2
import numpy as np
def dnn_inference_pipeline(net, img, input_name=None):
"""DNN 推理标准流程"""
# 1. 预处理：blobFromImage
blob = cv2.dnn.blobFromImage(
img,
scalefactor=1.0 / 255.0, # 归一化
size=(input_size, input_size), # 输入尺寸
mean=(123.675, 116.28, 103.53), # ImageNet 均值
swapRB=True, # BGR → RGB
crop=False
)
# 不同模型需要不同的预处理（需查模型文档）
# 2. 设置输入
if input_name:
net.setInput(blob, input_name)
else:
net.setInput(blob)
# 3. 前向推理
output = net.forward() # 默认输出第一个输出层
# 4. 后处理（根据模型类型自定义）
result = postprocess(output)
return result24.1.2 blobFromImage 参数详解
python
blob = cv2.dnn.blobFromImage(
image, # 输入图像
scalefactor=1.0/255.0, # 缩放因子（归一化到 [0,1]）
size=(w, h), # 目标输入尺寸
mean=(0, 0, 0), # 减均值（通常用 ImageNet 均值）
swapRB=False, # True=BGR→RGB, False=不交换
crop=False, # 中心裁剪
ddepth=cv2.CV_32F # 输出深度（必须 float32）
)
# 输出形状: (1, C, H, W) — batch=124.2 图像分类
24.2.1 经典分类模型推理
python
import cv2
# ---- 加载分类模型 ----
net = cv2.dnn.readNetFromONNX('mobilenet_v3.onnx')
# ---- 推理 ----
img = cv2.imread('test.jpg')
blob = cv2.dnn.blobFromImage(img, 1.0/255.0, (224, 224),
mean=(123.675, 116.28, 103.53),
swapRB=True, crop=False)
net.setInput(blob)
output = net.forward() # (1, 1000) — ImageNet 1000 类概率
# ---- 取 top-5 ----
top5 = np.argsort(output.flatten())[-5:][::-1]
probabilities = output.flatten()[top5]
# ImageNet class labels (1000 classes)
with open('imagenet_classes.txt') as f:
classes = f.read().strip().split('\n')
for cls_id, prob in zip(top5, probabilities):
print(f'{classes[cls_id]}: {prob:.4f}')24.2.2 常用预训练分类模型
| 模型 | 输入尺寸 | 精度 | 速度 | 适用场景 |
|---|---|---|---|---|
| MobileNet V3 | 224×224 | ★★★ | ⚡⚡⚡ | 移动端 ✅ |
| EfficientNet B0 | 224×224 | ★★★★ | ⚡⚡ | 平衡 ✅ |
| ResNet-50 | 224×224 | ★★ | ⚡ | 高精度 |
| VGG-16 | 224×224 | ★★★★ | ⚡ | 经典 |
| NASNet-Mobile | 331×331 | ★★ | ⚡⚡ | 移动端 |
24.3 语义分割
24.3.1 DeepLab / PSPNet 推理
python
import cv2
import numpy as np
# DeepLab V3+ 语义分割
net = cv2.dnn.readNetFromONNX('deeplabv3.onnx')
img = cv2.imread('test.jpg')
h, w = img.shape[:2]
blob = cv2.dnn.blobFromImage(
img, 1.0/255.0, (513, 513),
mean=(123.675, 116.28, 103.53),
swapRB=True,
crop=False
)
net.setInput(blob)
output = net.forward() # (1, N, 513, 513) — N 个类别
# 后处理：argmax → 语义标签图
seg_map = np.argmax(output[0], axis=0) # (513, 513)
# 上采样到原图尺寸
seg_map_resized = cv2.resize(seg_map, (w, h), interpolation=cv2.INTER_NEAREST)
# 着色可视化
color_seg = apply_colormap(seg_map_resized) # 自定义 colormap
overlay = cv2.addWeighted(img, 0.5, color_seg, 0.5, 0)24.3.2 常用 Colormap
python
def apply_colormap(sem_seg):
"""将语义分割图转为彩色图"""
h, w = sem_seg.shape
colored = np.zeros((h, w, 3), dtype=np.uint8)
color_palette = get_color_palette() # 21×3 的调色板
for i in range(21):
mask = sem_seg == i
colored[mask] = color_palette[i]
return colored
# Cityscapes 常用配色
def get_color_palette():
# 21 类 Cityscapes colormap
palette = [
[128, 64, 128], [244, 35, 232], [70, 70, 70], [102, 102, 156],
[190, 153, 153], [180, 165, 180], [150, 100, 100], [150, 120, 90],
[153, 153, 153], [153, 153, 153], [250, 170, 30], [220, 220, 0],
[107, 142, 35], [152, 251, 152], [70, 130, 180], [220, 20, 60],
[255, 0, 0], [0, 0, 142], [0, 0, 70], [0, 60, 100], [0, 80, 100]
]
return np.array(palette)24.4 人脸识别（OpenCV 内置）
python
import cv2
# ---- 加载预训练模型 ----
face_detector = cv2.dnn.readNetFromTorch('res10_300x300_ssd_iter_140000.caffemodel',
'deploy.prototxt')
# 或 ONNX 格式
# face_detector = cv2.dnn.readNetFromONNX('face_detection.onnx')
# ---- 人脸检测 ----
img = cv2.imread('test.jpg')
blob = cv2.dnn.blobFromImage(img, 1.0/255.0, (300, 300), mean=(104, 177, 123))
face_detector.setInput(blob)
detections = face_detector.forward() # (1, 1, N, 7)
for i in range(detections.shape[2]):
confidence = detections[0, 0, i, 2]
if confidence > 0.5:
x1 = int(detections[0, 0, i, 3] * img.shape[1])
y1 = int(detections[0, 0, i, 4] * img.shape[0])
x2 = int(detections[0, 0, i, 5] * img.shape[1])
y2 = int(detections[0, 0, i, 6] * img.shape[0])
cv2.rectangle(img, (x1, y1), (x2, y2), (0, 255, 0), 2)24.4.1 人脸对齐 + 识别
python
import cv2
import numpy as np
# 1. 人脸检测
faces = detect_faces(img)
for face in faces:
x, y, w, h = face
# 2. 面部关键点检测（5 点或 68 点）
keypoints = detect_landmarks(face_roi) # (5, 2) 或 (68, 2)
# 3. 仿射变换对齐
# 标准面部关键点（用于对齐）
src_pts = keypoints
dst_pts = np.array([ # 标准位置
[w*0.3, h*0.38],
[w*0.7, h*0.38],
[w*0.5, h*0.7],
[w*0.55, h*0.85],
[w*0.45, h*0.85],
], dtype=np.float32)
M = cv2.getAffineTransform(src_pts, dst_pts)
aligned = cv2.warpAffine(face_roi, M, (112, 112)) # 对齐到 112×112
# 4. 人脸识别（与 face_recognition 库或 DNN 模型结合）
# emb = face_recognition_model.forward(aligned)24.5 姿态估计（Pose Estimation）
python
# OpenPose / HRNet 风格姿态估计
net = cv2.dnn.readNetFromONNX('pose.onnx')
img = cv2.imread('test.jpg')
blob = cv2.dnn.blobFromImage(img, 1.0/255.0, (368, 368), mean=(0, 0, 0), swapRB=True)
net.setInput(blob)
output = net.forward() # (1, N, H/4, W/4) — N 个人体部位的概率图
# 后处理：argmax → 关键点位置
N = output.shape[1]
H, W = output.shape[2], output.shape[3]
keypoints = np.zeros((N, 2), dtype=np.int32)
for n in range(N):
prob_map = output[0, n]
max_idx = np.unravel_index(np.argmax(prob_map), prob_map.shape)
keypoints[n] = [max_idx[1] * 4, max_idx[0] * 4] # 上采样
# 人体部位连接
body_parts = [
('nose', 'left_eye'), ('left_eye', 'right_eye'),
('left_eye', 'left_shoulder'), ('right_eye', 'right_shoulder'),
('left_shoulder', 'right_shoulder'),
('left_shoulder', 'left_elbow'), ('right_shoulder', 'right_elbow'),
('left_elbow', 'left_wrist'), ('right_elbow', 'right_wrist'),
('left_shoulder', 'left_hip'), ('right_shoulder', 'right_hip'),
('left_hip', 'right_hip'), ('left_hip', 'left_knee'),
('right_hip', 'right_knee'), ('left_knee', 'left_ankle'),
('right_knee', 'right_ankle')
]
img_pose = img.copy()
for part1, part2 in body_parts:
idx1 = get_keypoint_index(part1)
idx2 = get_keypoint_index(part2)
p1 = keypoints[idx1]
p2 = keypoints[idx2]
if p1[0] > 0 and p2[0] > 0:
cv2.line(img_pose, (int(p1[0]), int(p1[1])),
(int(p2[0]), int(p2[1])), (0, 255, 0), 2)24.6 文字检测（OCR 基础）
24.6.1 EAST 文字检测
python
# EAST 文字检测器
net = cv2.dnn.readNetFromTorch('frozen_east_text_detection.pb')
img = cv2.imread('test.jpg')
h, w = img.shape[:2]
blob = cv2.dnn.blobFromImage(img, 1.0, (320, 320), mean=(123.68, 116.78, 103.94),
swapRB=True, crop=False)
net.setInput(blob)
scores, geometry = net.forward(['feature_fusion/Decision_934', 'feature_fusion/Conv_7/Sigmoid'])
# 后处理：非极大值抑制检测文字框
boxes, confidences = detect_text_boxes(scores, geometry, w, h, conf_threshold=0.5)