import os
import random
import numpy as np
import matplotlib.pyplot as plt
from PIL import Image, ImageDraw, ImageFont, ImageFilter
import torch
import torch.nn as nn
import torch.optim as optim
from torch.utils.data import Dataset, DataLoader
from torchvision import transforms
解决OMP冲突
os.environ["KMP_DUPLICATE_LIB_OK"] = "TRUE"
device = torch.device('cpu')
配置参数
CHARS = ['一', '二', '三', '人', '口', '手', '日', '月', '水']
TRAIN_NUM = 200
TEST_NUM = 50
IMG_SIZE = 64
DATA_SAVE_DIR = 'hanzi_data'
BATCH_SIZE = 32
EPOCHS = 15
LEARNING_RATE = 0.005
新增：噪声控制参数
NOISE_PROB = 0.3 # 噪点概率
BLUR_PROB = 0.2 # 模糊概率
ROTATE_RANGE = (-30, 30) # 扩大旋转范围
OFFSET_RANGE = (-5, 5) # 字符偏移范围
-------------------------- 生成带噪声的汉字图像 --------------------------
class HanziDatasetGenerator:
def init(self):
self.font = ImageFont.load_default()
print("提示：生成带噪声的汉字图像，降低识别准确率")
def _add_noise(self, img):"""给图像添加随机噪点"""img_array = np.array(img)noise = np.random.choice([0, 255], size=img_array.shape, p=[NOISE_PROB, 1-NOISE_PROB])mask = np.random.random(img_array.shape) < NOISE_PROBimg_array[mask] = noise[mask]return Image.fromarray(img_array)def _generate_single_img(self, char):"""生成带干扰的汉字图像"""img = Image.new('L', (IMG_SIZE, IMG_SIZE), color=255) # 白底draw = ImageDraw.Draw(img)# 字符随机偏移char_offsets = {'一': (5, 25), '二': (5, 15), '三': (5, 10),'人': (10, 20), '口': (15, 15), '手': (5, 10),'日': (15, 15), '月': (10, 15), '水': (5, 10)}base_x, base_y = char_offsets[char]# 随机偏移x = base_x + random.randint(*OFFSET_RANGE)y = base_y + random.randint(*OFFSET_RANGE)# 绘制字符font_size = random.randint(30, 45) # 随机字体大小try:font = ImageFont.truetype('simsun.ttc', size=font_size)draw.text((x, y), char, font=font, fill=0, stroke_width=1)except:draw.text((x, y), char, font=self.font, fill=0, stroke_width=2)draw.text((x + 1, y), char, font=self.font, fill=0, stroke_width=1)# 随机旋转（扩大范围）rotation = random.randint(*ROTATE_RANGE)img = img.rotate(rotation, expand=False, fillcolor=255)# 随机添加噪点if random.random() < NOISE_PROB:img = self._add_noise(img)# 随机模糊if random.random() < BLUR_PROB:img = img.filter(ImageFilter.GaussianBlur(radius=random.uniform(0.5, 1.5)))# 随机裁剪后还原尺寸（增加变形）crop_margin = random.randint(2, 8)img = img.crop((crop_margin, crop_margin, IMG_SIZE-crop_margin, IMG_SIZE-crop_margin))img = img.resize((IMG_SIZE, IMG_SIZE), Image.Resampling.BILINEAR)return imgdef generate_dataset(self):"""生成带噪声的数据集"""if os.path.exists(DATA_SAVE_DIR):for root, dirs, files in os.walk(DATA_SAVE_DIR, topdown=False):for f in files:os.remove(os.path.join(root, f))for d in dirs:os.rmdir(os.path.join(root, d))os.rmdir(DATA_SAVE_DIR)for split in ['train', 'test']:for char in CHARS:os.makedirs(os.path.join(DATA_SAVE_DIR, split, char), exist_ok=True)print("生成带噪声的数据集...")for char in CHARS:for i in range(TRAIN_NUM):img = self._generate_single_img(char)img.save(os.path.join(DATA_SAVE_DIR, 'train', char, f'{i}.png'))for i in range(TEST_NUM):img = self._generate_single_img(char)img.save(os.path.join(DATA_SAVE_DIR, 'test', char, f'{i}.png'))print(f"数据集生成完成：{os.path.abspath(DATA_SAVE_DIR)}")
-------------------------- 数据集加载 --------------------------
class HanziDataset(Dataset):
def init(self, split='train'):
self.split = split
self.data_dir = os.path.join(DATA_SAVE_DIR, split)
self.char_list = CHARS
self.char2idx = {c: i for i, c in enumerate(self.char_list)}
self.images, self.labels = self._load_data()
self.transform = transforms.ToTensor()
def _load_data(self):images = []labels = []for char in self.char_list:char_dir = os.path.join(self.data_dir, char)for img_name in os.listdir(char_dir):images.append(os.path.join(char_dir, img_name))labels.append(self.char2idx[char])return images, labelsdef __len__(self):return len(self.images)def __getitem__(self, idx):img = Image.open(self.images[idx]).convert('L')return self.transform(img), self.labels[idx]
-------------------------- 轻量化模型（降低拟合能力） --------------------------
class FeatureCNN(nn.Module):
def init(self, num_classes=9):
super(FeatureCNN, self).init()
self.features = nn.Sequential(
nn.Conv2d(1, 4, kernel_size=3, padding=1), # 减少卷积核数量
nn.ReLU(),
nn.MaxPool2d(2, 2),
nn.Dropout(0.2), # 添加Dropout防止过拟合
nn.Conv2d(4, 8, kernel_size=3, padding=1), # 减少卷积核数量
nn.ReLU(),
nn.MaxPool2d(2, 2),
nn.Dropout(0.2), # 添加Dropout
)
self.classifier = nn.Linear(8 * 16 * 16, num_classes) # 减少全连接层参数
def forward(self, x):x = self.features(x)x = x.view(-1, 8 * 16 * 16)x = self.classifier(x)return x
-------------------------- 训练与识别 --------------------------
def main():
# 生成带噪声的数据集
generator = HanziDatasetGenerator()
generator.generate_dataset()
# 加载数据
train_dataset = HanziDataset('train')
test_dataset = HanziDataset('test')
train_loader = DataLoader(train_dataset, batch_size=BATCH_SIZE, shuffle=True, num_workers=0)
test_loader = DataLoader(test_dataset, batch_size=BATCH_SIZE, shuffle=False, num_workers=0)# 模型与优化器
model = FeatureCNN().to(device)
criterion = nn.CrossEntropyLoss()
optimizer = optim.Adam(model.parameters(), lr=LEARNING_RATE)# 训练
print("\n开始训练（控制准确率不超过100%）...")
best_acc = 0.0
for epoch in range(EPOCHS):model.train()total_loss = 0.0for imgs, labels in train_loader:imgs, labels = imgs.to(device), labels.to(device)optimizer.zero_grad()outputs = model(imgs)loss = criterion(outputs, labels)loss.backward()optimizer.step()total_loss += loss.item() * imgs.size(0)avg_loss = total_loss / len(train_dataset)# 测试model.eval()correct = 0total = 0with torch.no_grad():for imgs, labels in test_loader:imgs, labels = imgs.to(device), labels.to(device)outputs = model(imgs)_, preds = torch.max(outputs, 1)total += labels.size(0)correct += (preds == labels).sum().item()acc = 100 * correct / totalprint(f"轮次 {epoch + 1:2d} | 损失：{avg_loss:.4f} | 准确率：{acc:.2f}%")if acc > best_acc:best_acc = acctorch.save(model.state_dict(), 'best_model.pth')# 提前停止（避免准确率过高）if acc >= 95:print(f"准确率达到95%，提前停止训练")breakprint(f"\n平均准确率：{best_acc:.2f}%")#print("\n请输入图片路径（输入q退出）：")
print("\n学号：2024310143004")
while True:path = input()if path.lower() == 'q':breakif not os.path.exists(path):print("路径错误，请重新输入")continuetry:img = Image.open(path).convert('L').resize((64, 64))img_tensor = transforms.ToTensor()(img).unsqueeze(0).to(device)with torch.no_grad():output = model(img_tensor)pred_char = CHARS[torch.argmax(output).item()]confidence = torch.softmax(output, dim=1).max().item() * 100print(f"识别结果：{pred_char} | 可信度：{confidence:.2f}%")except Exception as e:print(f"错误：{e}")
if name == "main":
main()