引言
验证码（CAPTCHA, Completely Automated Public Turing test to tell Computers and Humans Apart）广泛应用于网站注册、登录及重要操作中，用以防止自动化程序批量请求。然而，随着深度学习与计算机视觉技术的飞速发展，基于卷积神经网络（CNN）的验证码识别精度已经接近人类水平。本文将以 PyTorch 框架为基础，完整介绍一个从数据生成到模型优化的验证码识别系统实现过程。
技术背景
验证码特征
字符集：数字（0-9）、字母（a-z/A-Z）或混合
干扰方式：曲线、噪点、字符扭曲
更多内容访问ttocr.com或联系1436423940
长度固定或可变（常见 4~6 位）
深度学习优势
能自动学习特征，减少人工特征工程
CNN 在图像分类与序列特征提取方面效果突出
可结合 CTC（Connectionist Temporal Classification）处理变长验证码
系统实现步骤
3.1 数据集生成
验证码识别模型对数据量要求较高，因此人工采集成本过大时，可利用 captcha 库自动生成。
from captcha.image import ImageCaptcha
import random, string, os
参数配置
chars = string.digits + string.ascii_lowercase
dataset_dir = "dataset"
os.makedirs(dataset_dir, exist_ok=True)
def random_text(length=4):
return ''.join(random.choice(chars) for _ in range(length))
def generate_dataset(n=5000):
gen = ImageCaptcha(width=160, height=60)
for i in range(n):
text = random_text()
img = gen.generate_image(text)
img.save(f"{dataset_dir}/{text}_{i}.png")
generate_dataset()
生成的文件命名方式为 {标签}_{编号}.png，方便后续解析。
3.2 数据加载与预处理
import torch
from torch.utils.data import Dataset, DataLoader
from PIL import Image
import glob
class CaptchaDataset(Dataset):
def init(self, path, transform=None):
self.files = glob.glob(f"{path}/*.png")
self.transform = transform
self.chars = chars
self.char_to_idx = {c: i for i, c in enumerate(self.chars)}
def len(self):
return len(self.files)
def getitem(self, idx):
file = self.files[idx]
label_str = file.split("/")[-1].split("_")[0]
label = torch.tensor([self.char_to_idx[c] for c in label_str], dtype=torch.long)
img = Image.open(file).convert("L")
if self.transform:
img = self.transform(img)
return img, label
数据增强与归一化
from torchvision import transforms
transform = transforms.Compose([
transforms.Resize((60, 160)),
transforms.ToTensor()
])
dataset = CaptchaDataset("dataset", transform=transform)
loader = DataLoader(dataset, batch_size=64, shuffle=True)
3.3 模型设计
采用卷积层提取特征 + 全连接层进行多字符预测。
import torch.nn as nn
class CaptchaCNN(nn.Module):
def init(self, num_classes=len(chars), captcha_length=4):
super().init()
self.captcha_length = captcha_length
self.conv = nn.Sequential(
nn.Conv2d(1, 32, 3, padding=1), nn.ReLU(), nn.MaxPool2d(2),
nn.Conv2d(32, 64, 3, padding=1), nn.ReLU(), nn.MaxPool2d(2),
nn.Conv2d(64, 128, 3, padding=1), nn.ReLU(), nn.MaxPool2d(2)
)
self.fc = nn.Sequential(
nn.Linear(128 * 7 * 20, 1024),
nn.ReLU(),
nn.Dropout(0.25),
nn.Linear(1024, captcha_length * num_classes)
)
def forward(self, x):
x = self.conv(x)
x = x.view(x.size(0), -1)
x = self.fc(x)
return x.view(-1, self.captcha_length, len(chars))
3.4 模型训练
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model = CaptchaCNN().to(device)
criterion = nn.CrossEntropyLoss()
optimizer = torch.optim.Adam(model.parameters(), lr=1e-3)
for epoch in range(20):
model.train()
total_loss = 0
for imgs, labels in loader:
imgs, labels = imgs.to(device), labels.to(device)
optimizer.zero_grad()
output = model(imgs)
loss = sum(criterion(output[:, i, :], labels[:, i]) for i in range(4))
loss.backward()
optimizer.step()
total_loss += loss.item()
print(f"Epoch {epoch+1}, Loss: {total_loss / len(loader):.4f}")
3.5 测试与预测
def decode(pred):
pred_idx = pred.argmax(2)
return [''.join(chars[i] for i in p) for p in pred_idx]
model.eval()
test_imgs, test_labels = next(iter(loader))
with torch.no_grad():
preds = model(test_imgs.to(device))
pred_texts = decode(preds.cpu())
real_texts = [''.join(chars[i] for i in label) for label in test_labels]
for p, r in zip(pred_texts[:5], real_texts[:5]):
print(f"Pred: {p}, Real: {r}")