深度学习技术在计算机视觉领域的应用日益广泛，手写数字和符号的自动识别在教育、金融、邮政等领域具有重要意义。本博客将系统地介绍如何构建一个基于YOLO模型的手写数字和符号识别系统，涵盖数据准备、模型训练、UI界面开发和部署。
一、引言
项目背景与意义
手写数字和符号的自动识别系统可以大大提高信息处理的效率，减少人工录入的错误率。利用深度学习技术，尤其是YOLO（You Only Look Once）目标检测模型，可以实现对手写字符的高效、准确识别。
目标
- 构建一个基于YOLO的深度学习模型，能够准确识别手写数字和符号
- 开发一个用户友好的Web界面，方便用户上传图片并获取识别结果
- 部署系统，实现实时在线识别
二、技术方案
开发环境
- 操作系统：Windows/Linux/MacOS
- 编程语言：Python 3.8+
- 开发工具：PyCharm/VSCode
- 深度学习框架：PyTorch
- Web框架：Flask
依赖库安装
首先，创建一个新的Python虚拟环境并安装所需的依赖库：
bash
conda create -n handwriting_recognition python=3.8
conda activate handwriting_recognition
pip install torch torchvision torchaudio
pip install flask opencv-python pandas
pip install -U git+https://github.com/ultralytics/yolov5三、数据准备
数据集获取
使用公开的手写数字和符号数据集，如MNIST和Kuzushiji-MNIST。可以从以下链接下载：
- MNIST: http://yann.lecun.com/exdb/mnist/
- Kuzushiji-MNIST: https://github.com/rois-codh/kmnist
数据预处理
将数据集转换为YOLO格式，需要包含图像文件和相应的标签文件。标签文件格式如下：
<class_id> <x_center> <y_center> <width> <height>
编写数据预处理脚本，将数据集转换为YOLO格式：
python
import os
import pandas as pd
from PIL import Image
def convert_to_yolo_format(df, output_dir):
for index, row in df.iterrows():
img = Image.open(row['image_path'])
width, height = img.size
x_center = (row['x_min'] + row['x_max']) / 2 / width
y_center = (row['y_min'] + row['y_max']) / 2 / height
w = (row['x_max'] - row['x_min']) / width
h = (row['y_max'] - row['y_min']) / height
label_path = os.path.join(output_dir, f"{row['image_id']}.txt")
with open(label_path, 'w') as f:
f.write(f"{row['class_id']} {x_center} {y_center} {w} {h}\n")
# Example usage
# df = pd.read_csv('annotations.csv') # Assuming annotations.csv has columns: image_id, image_path, x_min, y_min, x_max, y_max, class_id
# convert_to_yolo_format(df, 'labels/')数据集划分
将数据集划分为训练集、验证集和测试集：
python
import shutil
import random
def split_dataset(source_dir, train_dir, val_dir, test_dir, train_ratio=0.7, val_ratio=0.2):
all_files = os.listdir(source_dir)
random.shuffle(all_files)
train_count = int(len(all_files) * train_ratio)
val_count = int(len(all_files) * val_ratio)
for i, file in enumerate(all_files):
if i < train_count:
shutil.move(os.path.join(source_dir, file), train_dir)
elif i < train_count + val_count:
shutil.move(os.path.join(source_dir, file), val_dir)
else:
shutil.move(os.path.join(source_dir, file), test_dir)
split_dataset('data/images', 'data/train/images', 'data/val/images', 'data/test/images')
split_dataset('data/labels', 'data/train/labels', 'data/val/labels', 'data/test/labels')四、模型训练
配置YOLO模型
下载YOLOv5预训练权重，并配置数据文件：
yaml
# handwrite.yaml
train: data/train
val: data/val
nc: 10 # number of classes (0-9 for MNIST, adjust accordingly for symbols)
names: ['0', '1', '2', '3', '4', '5', '6', '7', '8', '9']模型训练
运行以下命令开始训练：
bash
python train.py --img 640 --batch 16 --epochs 50 --data handwrite.yaml --cfg yolov5s.yaml --weights yolov5s.pt模型评估
使用验证集评估模型性能，并进行必要的超参数调优：
python
from sklearn.metrics import accuracy_score, recall_score, f1_score
y_true = [...] # true labels
y_pred = [...] # predicted labels
accuracy = accuracy_score(y_true, y_pred)
recall = recall_score(y_true, y_pred, average='macro')
f1 = f1_score(y_true, y_pred, average='macro')
print(f"Accuracy: {accuracy}, Recall: {recall}, F1 Score: {f1}")五、UI界面开发
Flask搭建Web应用
-
创建项目目录结构：
handwriting_recognition/ ├── app.py ├── templates/ │ ├── index.html │ └── result.html ├── static/ │ └── uploads/ └── models/ └── yolov5s.pt -
编写网页模板：
-
index.html
html<!DOCTYPE html> <html lang="en"> <head> <meta charset="UTF-8"> <meta name="viewport" content="width=device-width, initial-scale=1.0"> <title>Handwriting Recognition</title> <link rel="stylesheet" href="{{ url_for('static', filename='styles.css') }}"> </head> <body> <h1>Handwriting Recognition</h1> <form action="/predict" method="post" enctype="multipart/form-data"> <input type="file" name="file"> <button type="submit">Upload</button> </form> </body> </html> -
result.html
html<!DOCTYPE html> <html lang="en"> <head> <meta charset="UTF-8"> <meta name="viewport" content="width=device-width, initial-scale=1.0"> <title>Result</title> <link rel="stylesheet" href="{{ url_for('static', filename='styles.css') }}"> </head> <body> <h1>Detection Result</h1> <img src="{{ url_for('static', filename='uploads/' + filename) }}" alt="Uploaded Image"> <p>{{ result }}</p> </body> </html>
-
实现后端逻辑
-
python
from flask import Flask, request, render_template, url_for import os from werkzeug.utils import secure_filename import torch from PIL import Image app = Flask(__name__) app.config['UPLOAD_FOLDER'] = 'static/uploads/' model = torch.hub.load('ultralytics/yolov5', 'custom', path='models/yolov5s.pt') @app.route('/') def index(): return render_template('index.html') @app.route('/predict', methods=['POST']) def predict(): if 'file' not in request.files: return 'No file part' file = request.files['file'] if file.filename == '': return 'No selected file' if file: filename = secure_filename(file.filename) filepath = os.path.join(app.config['UPLOAD_FOLDER'], filename) file.save(filepath) img = Image.open(filepath) results = model(img) results.save(save_dir=app.config['UPLOAD_FOLDER']) return render_template('result.html', filename=filename, result=results.pandas().xyxy[0].to_json(orient="records")) if __name__ == '__main__': app.run(debug=True)
六、模型部署
部署到云服务器
-
使用Gunicorn部署
bashpip install gunicorn gunicorn -w 4 app
:app
2. **配置Nginx反向代理**
```nginx
server {
listen 80;
server_name your_domain;
location / {
proxy_pass http://127.0.0.1:8000;
proxy_set_header Host $host;
proxy_set_header X-Real-IP $remote_addr;
proxy_set_header X-Forwarded-For $proxy_add_x_forwarded_for;
proxy_set_header X-Forwarded-Proto $scheme;
}
}
七、声明
声明：本文只是简单的项目思路，如有部署的想法，想要（UI界面+YOLOv8/v7/v6/v5代码+训练数据集+视频教学）的可以联系作者.