FaceNet-Optimal-v2 人脸识别模型

一、模型概述

FaceNet-Optimal-v2 是一个基于深度学习的人脸识别模型，采用 FaceNet 架构 + ArcFace 损失函数 进行训练。该模型能够将人脸图像转换为 128 维的特征向量（embedding），通过计算向量间的余弦相似度实现人脸身份识别。

核心特点

二、模型介绍

2.1 技术架构

输入图像 (3×112×112)
    ↓
人脸检测 (MTCNN)
    ↓
人脸对齐与裁剪
    ↓
特征提取网络 (ResNet + ArcFace)
    ↓
128维特征向量
    ↓
余弦相似度计算
    ↓
身份识别结果

2.2 技术参数

2.3 模型文件说明

三、模型训练

3.1 训练环境

# 依赖安装
pip install torch torchvision facenet-pytorch opencv-python pillow numpy scikit-learn

3.2 数据准备

data/
├── train/
│   ├── person_001/
│   │   ├── img_001.jpg
│   │   ├── img_002.jpg
│   │   └── ...
│   ├── person_002/
│   │   └── ...
│   └── ...
└── val/
    └── ...

3.3 训练代码示例

import torch
from torch import nn
from torch.utils.data import DataLoader
from facenet_pytorch import InceptionResnetV1
from torchvision import transforms, datasets

# 数据预处理
transform = transforms.Compose([
    transforms.Resize((112, 112)),
    transforms.ToTensor(),
    transforms.Normalize(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5])
])

# 加载数据集
train_dataset = datasets.ImageFolder('data/train', transform=transform)
train_loader = DataLoader(train_dataset, batch_size=32, shuffle=True)

# 初始化模型
num_classes = len(train_dataset.classes)
model = InceptionResnetV1(pretrained='vggface2', num_classes=num_classes)

# 损失函数和优化器
criterion = nn.CrossEntropyLoss()
optimizer = torch.optim.Adam(model.parameters(), lr=0.001)

# 训练循环
model.train()
for epoch in range(50):
    for images, labels in train_loader:
        outputs = model(images)
        loss = criterion(outputs, labels)

        optimizer.zero_grad()
        loss.backward()
        optimizer.step()

    print(f"Epoch [{epoch+1}/50], Loss: {loss.item():.4f}")

# 保存模型
torch.save(model.state_dict(), 'best_model.pth')

3.4 微调训练（推荐）

import torch
from torch import nn
from facenet_pytorch import InceptionResnetV1

# 加载预训练权重
model = InceptionResnetV1(pretrained=None, num_classes=128)
model.load_state_dict(torch.load('best_model.pth'))

# 替换分类头适配自己的数据集
num_new_classes = 100  # 替换为你的类别数
model.last_linear = nn.Linear(128, num_new_classes)

# 冻结前面的层，只训练分类头
for param in model.parameters():
    param.requires_grad = False
model.last_linear.requires_grad_(True)

# 继续训练...

四、模型推理

4.1 环境准备

pip install torch facenet-pytorch opencv-python pillow

4.2 基础推理

import torch
from PIL import Image
from facenet_pytorch import MTCNN, InceptionResnetV1

# 初始化人脸检测器
mtcnn = MTCNN(image_size=112, device='cpu')

# 加载识别模型
model = InceptionResnetV1(pretrained=None, num_classes=128)
model.load_state_dict(torch.load('inference_model.pth'))
model.eval()

# 加载并预处理图像
img = Image.open('test.jpg')
face_tensor = mtcnn(img)

if face_tensor is not None:
    # 提取特征向量
    with torch.no_grad():
        embedding = model(face_tensor.unsqueeze(0))

    print(f"特征向量维度: {embedding.shape}")  # [1, 128]
    print(f"特征向量: {embedding[0][:5]}...")  # 显示前5个值
else:
    print("未检测到人脸")

4.3 人脸比对

import torch
from PIL import Image
from facenet_pytorch import MTCNN, InceptionResnetV1
import torch.nn.functional as F

# 初始化
mtcnn = MTCNN(image_size=112, device='cpu')
model = InceptionResnetV1(pretrained=None, num_classes=128)
model.load_state_dict(torch.load('inference_model.pth'))
model.eval()

def get_embedding(image_path):
    """获取图像的特征向量"""
    img = Image.open(image_path)
    face_tensor = mtcnn(img)
    if face_tensor is None:
        return None
    with torch.no_grad():
        embedding = model(face_tensor.unsqueeze(0))
    return embedding

def compare_faces(img1_path, img2_path, threshold=0.8):
    """比较两张人脸是否为同一人"""
    emb1 = get_embedding(img1_path)
    emb2 = get_embedding(img2_path)

    if emb1 is None or emb2 is None:
        return None, 0.0

    # 计算余弦相似度
    similarity = F.cosine_similarity(emb1, emb2).item()
    is_same = similarity > threshold

    return is_same, similarity

# 使用示例
is_same, similarity = compare_faces('person1.jpg', 'person2.jpg')
print(f"相似度: {similarity:.4f}")
print(f"是否同一人: {'是' if is_same else '否'}")

4.4 批量识别

import os
import torch
from PIL import Image
from facenet_pytorch import MTCNN, InceptionResnetV1
import numpy as np

class FaceRecognizer:
    def __init__(self, model_path, threshold=0.8):
        self.mtcnn = MTCNN(image_size=112, device='cpu')
        self.model = InceptionResnetV1(pretrained=None, num_classes=128)
        self.model.load_state_dict(torch.load(model_path))
        self.model.eval()
        self.threshold = threshold
        self.known_embeddings = []
        self.known_names = []

    def register(self, image_path, name):
        """注册人脸"""
        embedding = self._get_embedding(image_path)
        if embedding is not None:
            self.known_embeddings.append(embedding)
            self.known_names.append(name)
            print(f"✅ 已注册: {name}")
        else:
            print(f"❌ 注册失败: {name}")

    def recognize(self, image_path):
        """识别身份"""
        embedding = self._get_embedding(image_path)
        if embedding is None:
            return "未知", 0.0

        max_similarity = -1
        best_match = "未知"

        for known_emb, name in zip(self.known_embeddings, self.known_names):
            similarity = torch.nn.functional.cosine_similarity(
                embedding, known_emb
            ).item()
            if similarity > max_similarity:
                max_similarity = similarity
                best_match = name

        if max_similarity > self.threshold:
            return best_match, max_similarity
        return "未知", max_similarity

    def _get_embedding(self, image_path):
        """获取特征向量"""
        img = Image.open(image_path)
        face_tensor = self.mtcnn(img)
        if face_tensor is None:
            return None
        with torch.no_grad():
            embedding = self.model(face_tensor.unsqueeze(0))
        return embedding

# 使用示例
recognizer = FaceRecognizer('inference_model.pth')
recognizer.register('photos/zhangsan.jpg', '张三')
recognizer.register('photos/lisi.jpg', '李四')

result, confidence = recognizer.recognize('test.jpg')
print(f"识别结果: {result}, 置信度: {confidence:.4f}")

五、模型效果

5.1 训练指标

5.2 性能表现

5.3 推理速度

六、应用场景

• 🚪 门禁系统: 刷脸开门，安全便捷
• 🆔 身份验证: 人脸登录、支付认证
• 👥 人脸聚类: 自动分组相似人脸
• 🔍 人脸搜索: 从数据库中查找相似人脸
• 📸 智能相册: 自动整理人物照片

七、常见问题

Q1: 如何处理检测不到人脸的情况？

A: 确保图像中人脸清晰、光线充足，可调整 MTCNN 的 min_face_size 参数：

mtcnn = MTCNN(image_size=112, min_face_size=40)

Q2: 如何提高识别准确率？

A:

• 增加训练数据量（每人至少 20 张照片）
• 使用数据增强（翻转、旋转、亮度调整）
• 调整 ArcFace 的 margin 参数

Q3: 支持多人同时识别吗？

A: 支持，只需检测多人脸并逐一提取特征：

boxes, probs = mtcnn.detect(img, landmarks=False)
for box in boxes:
    face_tensor = mtcnn.extract(img, [box])
    # 处理每个检测到的人脸

八、许可证

Apache License 2.0

九、作者

Brilliant

十、相关链接

• GitHub 项目