• 🍨 本文为🔗365天深度学习训练营 中的学习记录博客
• 🍖 原作者：K同学啊

1. 简介 & 数据集介绍

本文的核心思路是解决非标准分类数据集（即无法简单通过文件夹分类）的识别问题，以车牌识别为例。
数据集中共包含 13,675 张车牌照片。

2. 环境

• 语言环境：Python 3.12.7
• 编译器：Jupyter Notebook
• 深度学习环境：torch—2.8.0 + cu126 / torchvision—0.23.1+cu126

3. 代码实现

3.1 前期准备

3.1.1 设置GPU & 导入库

from torchvision.transforms import transforms
from torch.utils.data import DataLoader,Dataset
import torch.utils.data as data
from torchvision import datasets
import torchvision.models as models
import torch.nn.functional as F
import torch.nn as nn
import torch,torchvision
import os,PIL,random,pathlib
import matplotlib.pyplot as plt
import numpy as np
import pandas as pd
from torchvision.io import read_image
from PIL import Image
import torchsummary
from torch.autograd import Variable
from datetime import datetime

plt.rcParams['font.sans-serif'] = ['SimHei']
plt.rcParams['axes.unicode_minus'] = False

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
device

3.1.2 标签查看

这部分代码利用 pathlib 库遍历数据集目录。由于车牌的标签直接包含在文件名中（例如 000_川W9BR26.jpg），代码通过字符串分割操作提取出车牌号码。这是处理“无固定文件夹分类”数据的常用技巧。

data_dir = './Data/015_licence_plate/'
data_dir = pathlib.Path(data_dir)

data_paths  = list(data_dir.glob('*'))
classeNames = [str(path).split("\\")[2].split("_")[1].split(".")[0] for path in data_paths]
print(classeNames)

data_paths = list(data_dir.glob('*'))
data_paths_str = [str(path) for path in data_paths]
data_paths_str

3.1.3 数据可视化

使用 matplotlib 展示前18张原始车牌图像。这步操作的目的是验证数据读取是否正确，以及初步观察图像质量（如是否存在光照、倾斜等干扰），为后续的预处理（Resize、Normalize）提供感官依据。

plt.figure(figsize=(14,5))
plt.suptitle("数据示例",fontsize=15)

for i in range(18):
    plt.subplot(3,6,i+1)
    images = plt.imread(data_paths_str[i])
    plt.imshow(images)

plt.show()

3.1.4 标签数字化

这部分定义了字符集（省份简称+数字+字母共69类），并编写 text2vec 函数。该函数将 7 位车牌字符串转为 $7\times 69$ 的二进制矩阵（One-hot 形式）。这是将“文本识别”转化为“分类任务”的关键桥梁。

char_enum = ["京","沪","津","渝","冀","晋","蒙","辽","吉","黑","苏","浙","皖","闽","赣","鲁",\
              "豫","鄂","湘","粤","桂","琼","川","贵","云","藏","陕","甘","青","宁","新","军","使"]

number = [str(i) for i in range(0, 10)]    # 0 到 9 的数字
alphabet = [chr(i) for i in range(65, 91)]   # A 到 Z 的字母

char_set  = char_enum + number + alphabet
char_set_len = len(char_set)
label_name_len = len(classeNames[0])

# 将字符串数字化
def text2vec(text):
    vector = np.zeros([label_name_len, char_set_len])
    for i, c in enumerate(text):
        idx = char_set.index(c)
        vector[i][idx] = 1.0
    return vector

all_labels = [text2vec(i) for i in classeNames]

3.2 自建 CNN 模型

3.2.1 加载数据文件

自定义一个 MyDataset 加载车牌数据集。通过继承 torch.utils.data.Dataset，重写 getitem 方法，实现了在读取图像的同时同步加载预先编码好的数值标签。这种灵活的数据加载方式是处理 OCR 或目标检测等复杂任务的必备技能。

class MyDataset(data.Dataset):
    def __init__(self, all_labels, data_paths_str, transform):
        self.img_labels = all_labels
        self.img_dir = data_paths_str
        self.transform = transform

    def __len__(self):
        return len(self.img_labels)

    def __getitem__(self, index):
        image = Image.open(self.img_dir[index]).convert('RGB')
        label = self.img_labels[index]
        
        if self.transform:
            image = self.transform(image)
            
        return image, label

total_datadir = '/content/drive/MyDrive/Data/015_licence_plate/'

train_transforms = transforms.Compose([
    transforms.Resize([224, 224]),
    transforms.ToTensor(),
    transforms.Normalize(
        mean=[0.485, 0.456, 0.406], 
        std =[0.229, 0.224, 0.225])  # 其中 mean=[0.485,0.456,0.406]与std=[0.229,0.224,0.225] 从数据集中随机抽样计算得到的。
])

total_data = MyDataset(all_labels, data_paths_str, train_transforms)
total_data

3.2.2 数据集划分

利用 random_split 将总数据按 8:2 划分为训练集和测试集，并使用 DataLoader 封装。它负责在训练时按 batch_size=16 自动打乱数据并将其转化为模型可接受的 Tensor 类型，解决了内存限制与并行读取问题。

train_size = int(0.8 * len(total_data))
test_size  = len(total_data) - train_size
train_dataset, test_dataset = torch.utils.data.random_split(total_data, [train_size, test_size])
train_size,test_size

train_loader = torch.utils.data.DataLoader(train_dataset, batch_size=16, shuffle=True)
test_loader = torch.utils.data.DataLoader(test_dataset, batch_size=16, shuffle=True)

print("The number of images in a training set is: ", len(train_loader)*16)
print("The number of images in a test set is: ", len(test_loader)*16)
print("The number of batches per epoch is: ", len(train_loader))

for X, y in test_loader:
    print("Shape of X [N, C, H, W]: ", X.shape)
    print("Shape of y: ", y.shape, y.dtype)
    break

3.2.3 CNN 模型建立

这是一个标准的 CNN 模型，但特别加入了 BatchNorm2d 层以加速收敛。最关键的设计在于输出层：全连接层输出量设为 $7\times 69$，并通过自定义的 Reshape 层将数据重新排列为 [batch, 7, 69] 的形状，从而匹配车牌的 7 个位置。

class Network_bn(nn.Module):
    def __init__(self):
        super(Network_bn, self).__init__()
        """
        nn.Conv2d()函数：
        第一个参数（in_channels）是输入的channel数量
        第二个参数（out_channels）是输出的channel数量
        第三个参数（kernel_size）是卷积核大小
        第四个参数（stride）是步长，默认为1
        第五个参数（padding）是填充大小，默认为0
        """
        self.conv1 = nn.Conv2d(in_channels=3, out_channels=12, kernel_size=5, stride=1, padding=0)
        self.bn1 = nn.BatchNorm2d(12)
        self.conv2 = nn.Conv2d(in_channels=12, out_channels=12, kernel_size=5, stride=1, padding=0)
        self.bn2 = nn.BatchNorm2d(12)
        self.pool = nn.MaxPool2d(2,2)
        self.conv4 = nn.Conv2d(in_channels=12, out_channels=24, kernel_size=5, stride=1, padding=0)
        self.bn4 = nn.BatchNorm2d(24)
        self.conv5 = nn.Conv2d(in_channels=24, out_channels=24, kernel_size=5, stride=1, padding=0)
        self.bn5 = nn.BatchNorm2d(24)
        self.fc1 = nn.Linear(24*50*50, label_name_len*char_set_len)
        self.reshape = Reshape([label_name_len,char_set_len])

    def forward(self, x):
        x = F.relu(self.bn1(self.conv1(x)))
        x = F.relu(self.bn2(self.conv2(x)))
        x = self.pool(x)
        x = F.relu(self.bn4(self.conv4(x)))
        x = F.relu(self.bn5(self.conv5(x)))
        x = self.pool(x)
        x = x.view(-1, 24*50*50)
        x = self.fc1(x)

        # 最终reshape
        x = self.reshape(x)

        return x

# 定义Reshape层
class Reshape(nn.Module):
    def __init__(self, shape):
        super(Reshape, self).__init__()
        self.shape = shape

    def forward(self, x):
        return x.view(x.size(0), *self.shape)

device = "cuda" if torch.cuda.is_available() else "cpu"
print("Using {} device".format(device))

model = Network_bn().to(device)
model

torchsummary.summary(model, (3, 224, 224))

3.2.4 训练 & 测试函数

这是训练的核心逻辑。train 函数执行反向传播更新参数。test 函数中需要特别注意准确率（ACC）的计算：由于形状是 [batch, 7, 69]，需要对第 3 维做 argmax 获取预测字符，并判定 7个位置全对（使用 .all(1)）才算一个样本识别正确。

def test(model, test_loader, loss_model, device):
    size = len(test_loader.dataset)
    num_batches = len(test_loader)

    model.eval()
    test_loss, correct = 0, 0

    with torch.no_grad():
        for X, y in test_loader:
            X, y = X.to(device), y.to(device)
            pred = model(X) # pred shape: [batch, 7, 69]

            test_loss += loss_model(pred, y).item()

            predicted = pred.argmax(2)

            if y.dim() == 3:
                target = y.argmax(2)
            else:
                target = y

            correct += (predicted == target).all(1).type(torch.float).sum().item()

    test_loss /= num_batches
    accuracy = correct / size # 计算百分比

    print(f"Test Error: \n Accuracy: {(100*accuracy):>0.1f}%, Avg loss: {test_loss:>8f} \n")
    return correct, test_loss

def train(model, train_loader, loss_model, optimizer, device):
    model = model.to(device)
    model.train()

    for i, (images, labels) in enumerate(train_loader, 0):
        images = Variable(images.to(device))
        labels = Variable(labels.to(device))

        optimizer.zero_grad()
        outputs = model(images)

        loss = loss_model(outputs, labels)
        loss.backward()
        optimizer.step()

        if i % 1000 == 0:
            # 训练集 ACC 实时计算
            pred_classes = outputs.argmax(2)
            target_classes = labels.argmax(2) if labels.dim() == 3 else labels
            train_acc = (pred_classes == target_classes).all(1).type(torch.float).mean().item()

            print('[%5d] loss: %.3f | acc: %.3f' % (i, loss.item(), train_acc))

3.2.5 参数设定 & 正式训练

代码设定了 30 个 Epoch 的循环。每一轮迭代都会先在训练集上学习，再在测试集上评估。通过 test_acc_list 和 test_loss_list 记录整个进化过程，用于后续的收敛性能分析。

optimizer  = torch.optim.Adam(model.parameters(), lr=1e-4, weight_decay=0.0001)
loss_model = nn.CrossEntropyLoss()
test_acc_list  = []
test_loss_list = []
epochs = 30

for t in range(epochs):
    print(f"Epoch {t+1}\n-------------------------------")
    train(model, train_loader, loss_model, optimizer, device)
    test_acc,test_loss = test(model, test_loader, loss_model, device)
    test_acc_list.append(test_acc)
    test_loss_list.append(test_loss)
print("Done!")

3.2.6 结果可视化

使用 plt.plot 将训练过程中的 Loss 变化绘制成折线图。通过观察曲线是否平滑下降并趋于平稳，可以直观判断模型是否过拟合或收敛不足。带上时间戳（Current Time）则是为了确保实验记录的真实性和唯一性。

current_time = datetime.now()
x = [i for i in range(1,31)]

plt.plot(x, test_loss_list, label="Loss", alpha=0.8)
plt.xlabel("Epoch")
plt.ylabel("Loss")
plt.title(current_time)
plt.legend()
plt.show()