"""OpenOCR-MLT interactive demo using the published Python package.""" import logging import os from functools import cache from pathlib import Path from threading import RLock from time import perf_counter # Match the spaces SDK's accepted values and keep it optional for local CPU use. ZERO_GPU = os.environ.get("SPACES_ZERO_GPU", "").lower() in {"1", "t", "true"} if ZERO_GPU: import spaces import gradio as gr from PIL import ImageDraw, ImageFont from openocr_mlt import OpenOCRMLT ROOT = Path(__file__).resolve().parent DEVICE = os.environ.get("OPENOCR_MLT_DEVICE", "cuda:0" if ZERO_GPU else None) CACHE_DIR = os.environ.get("OPENOCR_MLT_CACHE_DIR") MODEL_REPOSITORY = "ericleyu/openocr-mlt-models" MODEL_REVISION = "7ef39160a893fc7e605e146c1544ba64c1a57f43" MODEL_ARCHIVE = "scriptmoe-medium-fp32-20261101-v1.tar.gz" # Display English labels while keeping the existing /predict task values. TASK_CHOICES = [ ("Whole-image OCR", "Whole-image OCR / 整图识别"), ("Text detection", "Text detection / 文字检测"), ("Cropped recognition", "Cropped text / 裁剪识别"), ] TASKS = {value: mode for (_, value), mode in zip(TASK_CHOICES, ("ocr", "det", "rec"))} LOCK = RLock() LOGGER = logging.getLogger("openocr-mlt-space") HEADER = """

OpenOCR-MLT

A visual demo for the openocr-mlt repository, developed by the FVL OpenOCR team.

Multilingual text recognition LatinChineseJapaneseKoreanArabic BanglaDevanagariCyrillicThaiTibetan

""" # Gradio 6.31 loads built-in controls in the browser locale before running head scripts. # Switch that locale through the same frontend function used by Gradio's settings. ENGLISH_LOCALE = r""" """ CSS = """ .app-header { text-align: center; max-width: 1200px; margin: 20px auto; padding: 24px 20px; } .app-header h1 { font-size: clamp(2rem, 5vw, 2.4rem); font-weight: 700; letter-spacing: -.01em; margin: 0 0 10px; } .app-header .tagline { font-size: 1.1em; line-height: 1.6; margin: 0 0 14px; } .app-header .script-line { font-size: .95em; line-height: 1.7; margin: 0; } .script-line strong { margin-right: 5px; } .script-line span { display: inline-block; background: color-mix(in srgb, currentColor 8%, transparent); border-radius: 999px; margin: 2px 3px; padding: 2px 10px; font-size: .88em; } .image-container img { object-fit: contain; } """ EXAMPLE_EXTENSIONS = {".png", ".jpg", ".jpeg", ".bmp", ".tif", ".tiff", ".webp"} def example_images(group): folder = ROOT / "examples" / group return [ str(path) for path in sorted(folder.glob("*")) if path.is_file() and path.suffix.lower() in EXAMPLE_EXTENSIONS ] @cache def get_engine(task): options = {} if task == "ocr": options = {"page_max_side": 2000, "rec_threshold": 0} elif task == "rec": options = {"max_side": 2000, "rec_threshold": 0} return OpenOCRMLT(task=task, device=DEVICE, cache_dir=CACHE_DIR, offline=ZERO_GPU, **options) def draw_boxes(image, boxes): canvas = image.copy() draw = ImageDraw.Draw(canvas) font = ImageFont.load_default(size=16) width = max(2, round(max(canvas.size) / 600)) for index, points in enumerate(boxes, 1): polygon = [tuple(point) for point in points] draw.line(polygon + [polygon[0]], fill="#059669", width=width) x = min(max(int(p[0]) for p in polygon), canvas.width - 1) y = min(max(0, min(int(p[1]) for p in polygon) - 22), canvas.height - 1) label = str(index) left, top, right, bottom = draw.textbbox((x, y), label, font=font) draw.rectangle((left - 2, top - 2, right + 3, bottom + 2), fill="#059669") draw.text((x, y), label, font=font, fill="white") return canvas def run_ocr(image, task_label, rec_threshold=0.5): if image is None: raise gr.Error("Upload an image or select an example.") task = TASKS.get(task_label) if task is None: raise gr.Error("Select an OCR task.") if image.width * image.height > 16_000_000: raise gr.Error("Please use an image with at most 16 megapixels.") if isinstance(rec_threshold, bool) or not isinstance(rec_threshold, (int, float)) or not 0 <= rec_threshold <= 1: raise gr.Error("Recognition confidence must be between 0 and 1.") threshold = float(rec_threshold) started = perf_counter() try: with LOCK: engine = get_engine(task) prediction = engine(image=image) device = engine.info["device"] except (OSError, RuntimeError, ValueError) as error: LOGGER.exception("OCR request failed for task=%s", task) raise gr.Error("Processing failed. Please retry or try another image.") from error elapsed = perf_counter() - started records = [item.to_dict() for item in (prediction.items if task == "ocr" else prediction)] score_key = "rec_score" if task == "ocr" else "score" table = [] for index, row in enumerate(records, 1): score = row[score_key] # Apply the slider to the display without recreating the model session. if row.get("text") and score < threshold: row.update(text="", status="filtered") table.append([index, row.get("text", ""), score, row.get("status", "detected")]) text = "\n".join(row["text"] for row in records if row.get("text")) boxes = [row["points"] for row in records if "points" in row] payload = { "task": task, "image_size": list(image.size), "device": device, "results": records, "recognition_threshold": None if task == "det" else threshold, } count = len(records) unit = "prediction" if task == "rec" else "region" status = f"{elapsed:.2f} s · {device.upper()} · {count} {unit}{'s' if count != 1 else ''}" if not records: status += " · No text found" elif task != "det" and not text: status += f" · No text passed confidence {threshold:g}" return draw_boxes(image, boxes), text, table, status, payload # ONNX Runtime opens its CUDA sessions on the first inference in the GPU worker. predict = spaces.GPU(duration=60)(run_ocr) if ZERO_GPU else run_ocr def create_demo(): cropped_examples = example_images("recognition") with gr.Blocks(title="OpenOCR-MLT") as demo: gr.HTML(HEADER) with gr.Row(): with gr.Column(): image = gr.Image( label="Input image", type="pil", sources=["upload"], elem_classes=["image-container"], height=380, ) task = gr.Radio( choices=TASK_CHOICES, value=TASK_CHOICES[0][1], label="Task", ) with gr.Column() as page_gallery: gr.Examples( examples=example_images("page"), inputs=image, label="Examples for whole-image OCR and text detection", cache_examples=False, ) with gr.Column(visible=False) as crop_gallery: if cropped_examples: gr.Examples( examples=cropped_examples, inputs=image, label="Cropped recognition examples", cache_examples=False, ) else: gr.Markdown("Upload an image cropped to a single word or line of text.") run = gr.Button("Run OCR", variant="primary") rec_threshold = gr.Slider( maximum=1, value=0.5, step=0.01, label="Recognition confidence", info="Minimum confidence for recognized text.", ) with gr.Column(): visualization = gr.Image( label="Visualization", elem_classes=["image-container"], height=380, ) status = gr.Textbox(label="Latency") structured = gr.JSON(label="Result (JSON)", height=310) with gr.Accordion("Recognized text and confidence scores", open=False): text = gr.Textbox(label="Recognized text", lines=6, buttons=["copy"]) table = gr.Dataframe( headers=["ID", "Text", "Confidence", "Status"], datatype=["number", "str", "number", "str"], wrap=True, label="Results", ) def update_task(label): mode = TASKS.get(label) return ( gr.update(visible=mode != "rec"), gr.update(visible=mode == "rec"), gr.update(visible=mode != "det"), gr.update(value={"det": "Run Detection", "rec": "Run Recognition"}.get(mode, "Run OCR")), ) task.change( update_task, inputs=task, outputs=[page_gallery, crop_gallery, rec_threshold, run], queue=False, api_visibility="private", ) run.click( predict, inputs=[image, task, rec_threshold], outputs=[visualization, text, table, status, structured], api_name="predict", concurrency_limit=1, ) gr.Markdown( "[PyPI](https://pypi.org/project/openocr-mlt/) · " "[Models](https://huggingface.co/ericleyu/openocr-mlt-models) · " "[ScriptMoE](https://github.com/YesianRohn/ScriptMoE) · " "[Reference demo and example images](https://huggingface.co/spaces/Yesianrohn/MultilingualOCR-Demo)" ) return demo.queue(max_size=12) def prepare_space_assets(): """Verify and unpack the build-preloaded model archive without opening CUDA sessions.""" from huggingface_hub import hf_hub_download from openocr_mlt import fetch_models archive = hf_hub_download( repo_id=MODEL_REPOSITORY, filename=MODEL_ARCHIVE, revision=MODEL_REVISION, local_files_only=True, ) fetch_models(source=archive, cache_dir=CACHE_DIR, offline=True) if __name__ == "__main__": if ZERO_GPU: prepare_space_assets() create_demo().launch( server_name="0.0.0.0", max_file_size="20mb", theme=gr.themes.Soft(), css=CSS, head=ENGLISH_LOCALE, )