extract_gtsrb.py

import os
import shutil
import random
from PIL import Image

import pandas as pd
import torch
import torchvision.transforms as T
from torchvision.transforms.functional import pad
from tqdm.auto import tqdm

# --- 配置参数 ---
INPUT_DIR = "./GTSRB"
OUTPUT_DIR = "./GTSRB_224_balanced"
IMAGE_SIZE = 224
TARGET_TRAIN_SAMPLES_PER_CLASS = 1000
RANDOM_STATE = 42  # 用于保证降采样的可复现性

# 设置随机种子以保证可复现性
torch.manual_seed(RANDOM_STATE)
random.seed(RANDOM_STATE)


# --- 自定义变换：保持比例填充到正方形 ---
class PadToSquare:
    def __init__(self, fill_value=0):
        self.fill_value = fill_value

    def __call__(self, img):
        w, h = img.size
        if w == h:
            return img
        max_dim = max(w, h)
        pad_left = (max_dim - w) // 2
        pad_right = max_dim - w - pad_left
        pad_top = (max_dim - h) // 2
        pad_bottom = max_dim - h - pad_top
        padding = (pad_left, pad_top, pad_right, pad_bottom)
        # 产生的空白区域用反射填充
        return pad(img, padding, padding_mode="reflect")


# --- 定义图像变换流程 ---
base_transform = T.Compose(
    [
        PadToSquare(fill_value=0),
        T.Resize(
            (IMAGE_SIZE, IMAGE_SIZE),
            interpolation=T.InterpolationMode.BICUBIC,
            antialias=True,
        ),
    ]
)

# - 关键：先执行基础变换放大图像，然后再在高分辨率图像上进行增强，防止锯齿放大。
augmentation_pipeline = T.Compose(
    [
        base_transform,
        # 产生的空白区域用中性的灰色填充
        T.RandomAffine(
            degrees=15, translate=(0.1, 0.1), scale=(0.9, 1.1), shear=5, fill=128
        ),
        T.ColorJitter(brightness=0.2, contrast=0.2, saturation=0.2),
    ]
)


def process_training_data():
    """处理训练数据集，包含降采样和数据增强逻辑。"""
    print("开始处理训练数据集...")

    train_input_dir = os.path.join(INPUT_DIR, "Final_Training", "Images")
    train_output_dir = os.path.join(OUTPUT_DIR, "train")

    class_dirs = [
        d
        for d in os.listdir(train_input_dir)
        if os.path.isdir(os.path.join(train_input_dir, d))
    ]

    for class_id_str in tqdm(class_dirs, desc="处理类别"):
        class_input_dir = os.path.join(train_input_dir, class_id_str)
        class_id_int = int(class_id_str)
        formatted_class_id = f"{class_id_int:02d}"
        class_output_dir = os.path.join(train_output_dir, formatted_class_id)
        os.makedirs(class_output_dir, exist_ok=True)

        annotation_file = os.path.join(class_input_dir, f"GT-{class_id_str}.csv")
        df = pd.read_csv(annotation_file, sep=";")

        # --- 平衡数据集逻辑 ---
        num_original = len(df)

        if num_original > TARGET_TRAIN_SAMPLES_PER_CLASS:
            # 降采样: 如果原始样本过多，随机选择1000个
            df_to_process = df.sample(
                n=TARGET_TRAIN_SAMPLES_PER_CLASS, random_state=RANDOM_STATE
            )
        else:
            # 否则，处理所有原始样本
            df_to_process = df

        # 1. 处理并保存最终要保留的原始图像
        images_for_augmentation_pool = []
        for _, row in df_to_process.iterrows():
            filename = row["Filename"]
            img_path = os.path.join(class_input_dir, filename)
            try:
                with Image.open(img_path) as img:
                    # 将用于增强的原始图存入内存
                    images_for_augmentation_pool.append(img.copy())
                    # 应用基础变换并保存
                    processed_img = base_transform(img)
                    output_filename = os.path.splitext(filename)[0] + ".jpeg"
                    processed_img.save(
                        os.path.join(class_output_dir, output_filename), "JPEG"
                    )
            except FileNotFoundError:
                print(f"警告：文件未找到 {img_path}")
            except Exception as e:
                print(f"处理文件 {img_path} 时出错: {e}")

        # 2. 如果样本数不足，进行数据增强
        num_processed = len(df_to_process)
        num_to_augment = TARGET_TRAIN_SAMPLES_PER_CLASS - num_processed

        if num_to_augment > 0:
            for i in range(num_to_augment):
                # 从我们已处理并保存的图片池中随机选择一张进行增强
                img_to_augment = random.choice(images_for_augmentation_pool)
                augmented_img = augmentation_pipeline(img_to_augment)
                output_filename = f"aug_{i:04d}.jpeg"
                augmented_img.save(
                    os.path.join(class_output_dir, output_filename), "JPEG"
                )

    print("训练数据集处理完成。")


def process_test_data():
    """处理测试数据集"""
    print("\n开始处理测试数据集...")
    test_input_dir = os.path.join(INPUT_DIR, "Final_Test", "Images")
    test_output_dir = os.path.join(OUTPUT_DIR, "test")
    annotation_file = os.path.join(INPUT_DIR, "GT-final_test.csv")
    df = pd.read_csv(annotation_file, sep=";")

    for _, row in tqdm(df.iterrows(), total=len(df), desc="处理测试图片"):
        filename = row["Filename"]
        class_id = row["ClassId"]
        img_path = os.path.join(test_input_dir, filename)
        formatted_class_id = f"{class_id:02d}"
        class_output_dir = os.path.join(test_output_dir, formatted_class_id)
        os.makedirs(class_output_dir, exist_ok=True)
        try:
            with Image.open(img_path) as img:
                processed_img = base_transform(img)
                output_filename = os.path.splitext(filename)[0] + ".jpeg"
                processed_img.save(
                    os.path.join(class_output_dir, output_filename), "JPEG"
                )
        except FileNotFoundError:
            print(f"警告：文件未找到 {img_path}")
        except Exception as e:
            print(f"处理文件 {img_path} 时出错: {e}")

    print("测试数据集处理完成。")


def main():
    print(f"GTSRB 数据集转换脚本")
    print(f"输入目录: {INPUT_DIR}")
    print(f"输出目录: {OUTPUT_DIR}\n")

    if os.path.exists(OUTPUT_DIR):
        print(f"发现已存在的输出目录 {OUTPUT_DIR}，正在删除...")
        shutil.rmtree(OUTPUT_DIR)

    print("正在创建新的输出目录结构...")
    os.makedirs(os.path.join(OUTPUT_DIR, "train"), exist_ok=True)
    os.makedirs(os.path.join(OUTPUT_DIR, "test"), exist_ok=True)

    process_training_data()
    process_test_data()

    print(f"\n🎉 全部处理完成！转换后的数据位于: {OUTPUT_DIR}")


if __name__ == "__main__":
    main()