让YOLOv11秒懂口罩/手套/防护服：CPPE5数据集实战指南

2025-08-06

目标检测的事儿，咱们不绕弯子了。这篇文章就是教你一件事：

3分钟教你用YOLOv11检测口罩、手套、防护服，工业、医疗、安防一网打尽。

这篇文章不会深究模型原理，而是手把手带你跑通全流程训练与推理，使用的数据是HuggingFace上经典的CPPE-5数据集，涵盖多种个人防护装备（PPE）目标检测任务。

✳️什么是 CPPE-5？

这里采用rishitdagli/cppe-5公开数据集，含以下5类标签：

类别英文	中文含义
Coverall	防护服 / 连体衣
Face_Shield	面罩 / 防护面屏
Gloves	手套
Goggles	护目镜
Mask	口罩

共1000张图片，真实复杂场景，适合目标检测实战测试。

🚀 快速开始（带你跑通）

1. 下载数据集

1
2
3

wget https://huggingface.co/datasets/rishitdagli/cppe-5/resolve/main/data/test-00000-of-00001.parquet

wget https://huggingface.co/datasets/rishitdagli/cppe-5/resolve/main/data/train-00000-of-00001.parquet

2. 转换成yolo格式

2.1 数据预处理脚本


import pathlib

import torch
from datasets import load_dataset
from PIL import Image, ImageDraw
from sympy import im

# 这里路径根据实际路径来
cppe5 = load_dataset(
    'parquet',
    data_files={
        "train": pathlib.Path(__file__).parent.parent.joinpath("train-00000-of-00001.parquet").__str__(),
        "test": pathlib.Path(__file__).parent.parent.joinpath("test-00000-of-00001.parquet").__str__()
    }
)
categories = cppe5["train"].features["objects"].feature["category"].names
id2label = {index: x for index, x in enumerate(categories, start=0)}
label2id = {v: k for k, v in id2label.items()}
# ['Coverall', 'Face_Shield', 'Gloves', 'Goggles', 'Mask']
print(categories)


def cppe5_to_yolo_format(examples, data_type='train'):
    # 注意这里，images/train放的是训练图片，images/test放的是测试图片
    # 同理，labels/train放的是对应图片标签
    pathlib.Path(__file__).parent.joinpath(
        f"images/{data_type}").mkdir(exist_ok=True, parents=True)
    pathlib.Path(__file__).parent.joinpath(
        f"labels/{data_type}").mkdir(exist_ok=True, parents=True)
    for index, sample in enumerate(examples):
        image = sample["image"].convert("RGB")
        yolo_labels = []

        # draw = ImageDraw.ImageDraw(image)
        # for bbox, category in zip(sample["objects"]["bbox"], sample["objects"]["category"]):
        #     x, y, w, h = tuple(bbox)
        #     if max(bbox) > 1.0:
        #         x1, y1 = int(x), int(y)
        #         x2, y2 = int(x + w), int(y + h)
        #     else:
        #         raise NotImplementedError
        #     draw.rectangle((x1, y1, x2, y2), outline="red")
        #     draw.text((x1, y1), str(id2label[category]), fill="red")
        # image.show()

        for bbox, category in zip(sample["objects"]["bbox"], sample["objects"]["category"]):

            x, y, w, h = tuple(bbox)
            if max(bbox) > 1.0:
                x1, y1 = int(x), int(y)
                x2, y2 = int(x + w), int(y + h)
            else:
                raise NotImplementedError
            img_w, img_h = image.size
            x_center = (x1 + x2) / 2.0 / img_w
            y_center = (y1 + y2) / 2.0 / img_h
            width = (x2 - x1) / img_w
            height = (y2 - y1) / img_h
            yolo_labels.append(
                f'{category} {x_center} {y_center} {width} {height}'
            )
        image.save(pathlib.Path(__file__).parent.joinpath(
            f"images/{data_type}/{index}.jpg"))
        with open(pathlib.Path(__file__).parent.joinpath(f"labels/{data_type}/{index}.txt"), "w") as f:
            f.write("\n".join(yolo_labels))


cppe5_train = cppe5_to_yolo_format(cppe5["train"], data_type='train')
cppe5_test = cppe5_to_yolo_format(cppe5["test"], data_type='test')

2.2 配置文件`data.yaml`

train: /object_detection/yolo_impl/images/train
val: /object_detection/yolo_impl/images/test
nc: 5
names: ["Coverall", "Face_Shield", "Gloves", "Goggles", "Mask"]

3. 启动YOLOv11模型训练


import os.path

from ultralytics import YOLO

os.environ['WANDB_DISABLED'] = 'true'
os.environ['YOLO_OFFLINE'] = 'true'


model_path = '/object_detection/yolo_impl/yolo11n.pt'
assert os.path.exists(model_path)

model = YOLO(model_path)

model.train(
    data=os.path.join(os.path.dirname(__file__), 'data.yaml'),
    epochs=100,             # 训练轮数
    imgsz=640,              # 输入图像尺寸
    batch=64,               # 批量大小，根据显存可调
    workers=4,              # 数据加载线程数

    # 优化器与学习率
    lr0=0.001,              # 初始学习率
    lrf=0.01,               # 最低学习率（cosine decay 最终值）
    optimizer='SGD',        # 优化器
    weight_decay=0.0005,    # 权重衰减
    warmup_epochs=3,        # 热身轮数
    patience=20,            # 提前停止的容忍度

    # 增强与正则
    augment=True,           # 启用数据增强
    mosaic=1.0,
    mixup=0.1,
    hsv_h=0.015,
    hsv_s=0.7,
    hsv_v=0.4,
    scale=0.5,
    translate=0.1,
    degrees=0.0,            # 不启用旋转

    # 损失函数超参
    box=7.5,                # 边界框损失权重
    cls=0.5,                # 类别损失权重

    iou=0.2,                # IoU 阈值（用于损失计算）

    device=0                # 使用GPU:0

)

📊 训练效果如下（示例）：

4. 推理预测


import pathlib

from ultralytics import YOLO

model = YOLO(pathlib.Path(__file__).parent.joinpath(
    "runs/detect/train/weights/best.pt"
    )
)

results = model.predict(

    source='/object_detection/yolo_impl/images/test',
    save=True

)
# 处理结果
for result in results:
    print(f"检测到的对象数量: {len(result.boxes)}")
    for box, conf, cls in zip(result.boxes.xyxy, result.boxes.conf, result.boxes.cls):
        print(f"类别: {result.names[int(cls)]}, 置信度: {conf:.2f}, 坐标: {box}")
        print(f"类别: {result.names[int(cls)]}, 置信度: {conf:.2f}, 坐标: {box}")