app.py

from PIL import Image
from ultralytics import YOLO

import os
import torch
import re
import cv2

import gradio as gr
import torchvision.transforms as T
import albumentations as A
import numpy as np
import matplotlib.pyplot as plt
import matplotlib.patches as patches

device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")

# Folders
input_folder = "./target"
output_folder = "./target_output"

os.makedirs(output_folder, exist_ok=True)

# Detector model
license_plate_detector = YOLO("./models/yolo11x.pt")

# SuperResolution model
sr = cv2.dnn_superres.DnnSuperResImpl_create()
sr.readModel("./models/FSRCNN_x3.pb")
sr.setModel("fsrcnn", 3)


class App:
    models = ['parseq', 'parseq_tiny', 'abinet', 'crnn', 'trba', 'vitstr']

    def __init__(self):
        self._model_cache = {}
        self._preprocess = T.Compose([
            T.Resize((32, 128), T.InterpolationMode.BICUBIC),
            T.ToTensor(),
            T.Normalize(0.5, 0.5)
        ])

    def _get_model(self, name):
        if name in self._model_cache:
            return self._model_cache[name]
        model = torch.hub.load('baudm/parseq', name, pretrained=True).eval().to(device)
        self._model_cache[name] = model
        return model

    @torch.inference_mode()
    def __call__(self, model_name, image):
        if image is None:
            return '', []
        model = self._get_model(model_name)
        image = self._preprocess(image.convert('RGB')).unsqueeze(0).to(device)
        pred = model(image).softmax(-1)
        label, _ = model.tokenizer.decode(pred)
        raw_label, raw_confidence = model.tokenizer.decode(pred, raw=True)
        max_len = 25 if model_name == 'crnn' else len(label[0]) + 1
        conf = list(map('{:0.1f}'.format, raw_confidence[0][:max_len].tolist()))
        return label[0], [raw_label[0][:max_len], conf]


p = App()

black_list = ["Y985BE152"]

def detect_license_plates(model, image):
    plate_image_np = pil_to_np(image)

    transform = A.Compose([
        A.ToGray(p=1.0),
        A.CLAHE(clip_limit=2.0, tile_grid_size=(8, 8), p=1.0),
    ])

    transformed = transform(image=plate_image_np)['image']

    if len(transformed.shape) == 2:
        transformed = cv2.cvtColor(transformed, cv2.COLOR_GRAY2RGB)

    image = np_to_pil(transformed)

    results = model(image)
    plates = []
    for result in results:
        for box in result.boxes.xyxy.cpu().numpy():
            x1, y1, x2, y2 = map(int, box)
            plate = image.crop((x1, y1, x2, y2))
            plates.append((plate, (x1, y1, x2, y2)))

    return plates


def pil_to_np(image):
    return np.array(image)


def np_to_pil(image_np):
    return Image.fromarray(image_np)


def preprocess_license_plate(plate_image: Image):
    plate_image_np = pil_to_np(plate_image)
    if not(plate_image_np.ndim == 2 or plate_image_np.shape[-1] == 1):
        plate_image_np = A.ToGray(p=1.0, num_output_channels=1)(image=plate_image_np)['image']
    super_resolved = sr.upsample(plate_image_np)
    augmented = A.Compose([
        A.CLAHE(clip_limit=2, tile_grid_size=(1, 1), p=1.0),
        A.Morphological(p=1.0, scale=(4, 4), operation="erosion"),
    ])(image=super_resolved)['image']

    super_resolved_pil = np_to_pil(augmented)
    return super_resolved_pil


def process_image(image_path: Image):
    image_np = np.array(image_path)

    fig, ax = plt.subplots(1, figsize=(10, 6))
    ax.imshow(image_np)

    plates = detect_license_plates(license_plate_detector, image_path)
    recognized_texts = []

    for i, (plate, bbox) in enumerate(plates):
        preprocessed_plate = preprocess_license_plate(plate)
        recognized_text, raw_output = p.__call__("parseq", preprocessed_plate)

        if recognized_text and len(recognized_text) > 5:
            recognized_text = re.sub(r"[^A-Za-z0-9]", "", recognized_text).upper()
            recognized_text = recognized_text.replace('V', 'Y').replace('I', '')
            recognized_text = recognized_text.replace('8', 'В', 1) if recognized_text[0] == "8" else recognized_text
            recognized_text = recognized_text.replace('7', 'T', 1) if recognized_text[0] == "7" else recognized_text
            recognized_text = recognized_text.replace('0', 'O', 1) if recognized_text[0] == "0" else recognized_text
            recognized_text = recognized_text[:9] if len(recognized_text) >= 9 else recognized_text
            if recognized_text not in black_list:
                recognized_texts.append(recognized_text)

        x1, y1, x2, y2 = bbox
        rect = patches.Rectangle((x1, y1), x2 - x1, y2 - y1, linewidth=2, edgecolor='r', facecolor='none')
        ax.add_patch(rect)
        ax.text(x1, y1 - 10, recognized_text, color='red', fontsize=12, bbox=dict(facecolor='white', alpha=0.5))

    plt.axis('off')

    # Saving image to buffer
    output_buffer = "processed_image.png"
    plt.savefig(output_buffer, bbox_inches='tight')
    plt.close()

    return Image.open(output_buffer), recognized_texts


# Gradio UI

target_folder = "./target"
example_images = [
    os.path.join(target_folder, file) for file in os.listdir(target_folder) if file.lower().endswith(("jpg", "png", "bmp"))
]

interface = gr.Interface(
    fn=process_image,
    inputs=gr.Image(type="pil", label="Загрузите фото машины с номером 📤"),
    outputs=[
        gr.Image(type="pil", label="📸 Выход 0 - Обработанное изображение"),
        gr.JSON(label="🔍 Выход 1 - Распознанный номер"),
    ],
    title="Распознавание российских номеров",
    description="🔎 **Загрузите изображение с автомобильным номером** и модель автоматически **определит госномер!** 🔥\n\n📸 **Форматы:** JPG, PNG, BMP",
    examples=example_images,
    flagging_mode="never",
    theme="compact",
)

if __name__ == "__main__":
    interface.launch(share=True)