app.py

"""
한국 번호판 OCR - KLPR_v1 (Model v5)
Hugging Face Gradio App
"""

import gradio as gr
import torch
import torch.nn as nn
from PIL import Image
import torchvision.transforms as transforms
import numpy as np

# ============================================================================
# 모델 정의
# ============================================================================
class CRNN(nn.Module):
    def __init__(self, img_height, num_chars, rnn_hidden=256):
        super(CRNN, self).__init__()

        # CNN - 32x200 -> 1x50
        self.cnn = nn.Sequential(
            nn.Conv2d(1, 64, kernel_size=3, padding=1),
            nn.ReLU(inplace=True),
            nn.MaxPool2d((2, 2)),

            nn.Conv2d(64, 128, kernel_size=3, padding=1),
            nn.ReLU(inplace=True),
            nn.MaxPool2d((2, 2)),

            nn.Conv2d(128, 256, kernel_size=3, padding=1),
            nn.BatchNorm2d(256),
            nn.ReLU(inplace=True),

            nn.Conv2d(256, 256, kernel_size=3, padding=1),
            nn.BatchNorm2d(256),
            nn.ReLU(inplace=True),
            nn.MaxPool2d((2, 1)),

            nn.Conv2d(256, 512, kernel_size=3, padding=1),
            nn.BatchNorm2d(512),
            nn.ReLU(inplace=True),

            nn.Conv2d(512, 512, kernel_size=3, padding=1),
            nn.BatchNorm2d(512),
            nn.ReLU(inplace=True),
            nn.MaxPool2d((2, 1)),

            nn.Conv2d(512, 512, kernel_size=3, padding=1),
            nn.BatchNorm2d(512),
            nn.ReLU(inplace=True),
            nn.MaxPool2d((2, 1))
        )

        self.rnn = nn.LSTM(512, rnn_hidden, bidirectional=True, num_layers=2, batch_first=True)
        self.fc = nn.Linear(rnn_hidden * 2, num_chars)

    def forward(self, x):
        conv = self.cnn(x)
        b, c, h, w = conv.size()
        conv = conv.squeeze(2).permute(0, 2, 1)
        rnn_out, _ = self.rnn(conv)
        output = self.fc(rnn_out)
        return output

# ============================================================================
# CTC 디코딩
# ============================================================================
def decode_predictions(outputs, itos, blank_idx=0):
    """CTC 디코딩"""
    preds = outputs.argmax(2).detach().cpu().numpy()  # (B, T)

    decoded = []
    for pred in preds:
        char_list = []
        prev_idx = blank_idx
        for idx in pred:
            if idx != blank_idx and idx != prev_idx:
                char_list.append(itos[int(idx)])
            prev_idx = idx
        decoded.append(''.join(char_list))
    return decoded

# ============================================================================
# 이미지 전처리
# ============================================================================
def preprocess_image(image, img_height=32, max_width=200):
    """번호판 이미지 전처리"""
    # PIL Image로 변환 (Gradio 4.x에서 type="pil"로 이미 PIL Image)
    if not isinstance(image, Image.Image):
        if isinstance(image, np.ndarray):
            image = Image.fromarray(image.astype('uint8'))

    image = image.convert('L')

    # 리사이즈 (aspect ratio 유지)
    w, h = image.size
    new_w = min(int(img_height * w / h), max_width)
    image = image.resize((new_w, img_height), Image.LANCZOS)

    # 패딩
    new_img = Image.new('L', (max_width, img_height), 255)
    new_img.paste(image, (0, 0))

    # Transform
    transform = transforms.Compose([
        transforms.ToTensor(),
        transforms.Normalize((0.5,), (0.5,))
    ])

    return transform(new_img).unsqueeze(0)  # (1, 1, H, W)

# ============================================================================
# 모델 로드
# ============================================================================
print("모델 로딩 중...")
checkpoint_path = 'best_ocr_one_line.pth'
checkpoint = torch.load(checkpoint_path, map_location='cpu')

img_h = checkpoint.get('img_h', 32)
max_w = checkpoint.get('max_w', 200)
itos = checkpoint['itos']
num_chars = len(itos)

device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
model = CRNN(img_h, num_chars, rnn_hidden=256).to(device)
model.load_state_dict(checkpoint['model_state'])
model.eval()

print(f"✓ 모델 로드 완료 (Device: {device})")
print(f"  - Epoch: {checkpoint.get('epoch', '?')}")
print(f"  - Val Acc: {checkpoint.get('val_acc', '?'):.2%}")

# ============================================================================
# 추론 함수
# ============================================================================
def predict_license_plate(image):
    """번호판 이미지에서 텍스트 예측"""
    if image is None:
        return "이미지를 업로드해주세요."

    try:
        # 전처리
        image_tensor = preprocess_image(image, img_h, max_w).to(device)

        # 추론
        with torch.no_grad():
            outputs = model(image_tensor).log_softmax(2)
            predictions = decode_predictions(outputs, itos)

        result = predictions[0]
        return result if result else "(인식 결과 없음)"

    except Exception as e:
        return f"오류 발생: {str(e)}"

# ============================================================================
# Gradio 인터페이스
# ============================================================================
demo = gr.Interface(
    fn=predict_license_plate,
    inputs=gr.Image(type="pil", label="번호판 이미지"),
    outputs=gr.Textbox(label="인식 결과"),
    title="🚗 한국 번호판 OCR - KLPR v2",
    description="""
    한국 자동차 번호판을 인식하는 OCR 모델입니다.

    **모델 정보:**
    - Model: CRNN (CNN + Bidirectional LSTM + CTC)
    - Validation Accuracy: 91.23%
    - Epoch: 18
    - 지원 문자: 77개 (한글 + 숫자 + 추가 특수 지역명)

    **사용 방법:**
    1. 번호판 이미지를 업로드하세요
    2. 자동으로 번호판 번호가 인식됩니다
    """,
    api_name="predict"
)

if __name__ == "__main__":
    demo.launch()