app.py

import io
import time
import os
import re
import numpy as np
from PIL import Image, ImageFilter
from cairosvg import svg2png
from transformers import VisionEncoderDecoderModel, TrOCRProcessor
import gradio as gr

processor = TrOCRProcessor.from_pretrained("anuashok/ocr-captcha-v3")
model = VisionEncoderDecoderModel.from_pretrained("anuashok/ocr-captcha-v3")
os.makedirs("outputs", exist_ok=True)


def _single_ocr_from_image(image: Image.Image) -> str:
    pixel_values = processor(image, return_tensors="pt").pixel_values
    generated_ids = model.generate(pixel_values)
    generated_text = processor.batch_decode(generated_ids, skip_special_tokens=True)[0]
    sanitized = re.sub(r'[^A-Z0-9]', '', generated_text).upper()
    return sanitized[:4]


def solve_svg_captcha(svg_data: str) -> str:
    svg = svg_data or ""
    svg = re.sub(r'<style>.*?</style>', '', svg, flags=re.DOTALL)
    svg = svg.replace('file:///', '')
    svg = svg.replace('/app/', '')
    svg = re.sub(r'url\(["\']?\/?app\/[^)"\']*["\']?\)', 'url()', svg)

    svg_static = re.sub(r'<animateTransform\b[^>]*>(?:.*?</animateTransform>)?', '', svg, flags=re.DOTALL)

    rotate_re = re.compile(r'rotate\(\s*([+-]?\d+)\s*,\s*([0-9.]+)\s*,\s*([0-9.]+)\s*\)')
    matches = rotate_re.findall(svg_static)
    centers = []
    seen = set()
    for _, cx, cy in matches:
        key = f"{cx},{cy}"
        if key not in seen:
            seen.add(key)
            centers.append((cx, cy))

    if not centers:
        try:
            png_bytes = svg2png(bytestring=svg_static.encode('utf-8'))
            image = Image.open(io.BytesIO(png_bytes)).convert("RGBA")
            image = image.resize((600, 400))
            background = Image.new("RGBA", image.size, (255, 255, 255))
            combined = Image.alpha_composite(background, image).convert("RGB")
            return _single_ocr_from_image(combined)
        except Exception as e:
            print("OCR fallback error:", e)
            return ""

    centers = centers[:2]
    angle_step = 15
    top_k = 2
    best_angles = {}

    for cx, cy in centers:
        metrics = []
        for angle in range(0, 360, angle_step):
            try:
                tmp = re.sub(rf'rotate\(\s*1\s*,\s*{re.escape(cx)}\s*,\s*{re.escape(cy)}\s*\)', f'rotate({angle}, {cx}, {cy})', svg_static)
                tmp = re.sub(rf'rotate\(\s*-1\s*,\s*{re.escape(cx)}\s*,\s*{re.escape(cy)}\s*\)', f'rotate(-{angle}, {cx}, {cy})', tmp)
                png_bytes = svg2png(bytestring=tmp.encode('utf-8'))
                img = Image.open(io.BytesIO(png_bytes)).convert('L')
                img = img.resize((600, 400))
                img = img.filter(ImageFilter.GaussianBlur(radius=1))
                edges = img.filter(ImageFilter.FIND_EDGES)
                arr = np.array(edges)
                edge_count = int((arr > 10).sum())
                metrics.append((edge_count, angle))
            except Exception:
                continue
        metrics.sort(key=lambda x: x[0])
        picked = [m[1] for m in metrics[:top_k]] if metrics else [0] * top_k
        if len(picked) < top_k:
            picked += [picked[0]] * (top_k - len(picked))
        best_angles[f"{cx},{cy}"] = picked

    combos = []
    if len(centers) == 1:
        k = f"{centers[0][0]},{centers[0][1]}"
        a1, a2 = best_angles[k][:2]
        combos = [{k: a1}, {k: a2}, {k: a1}, {k: a2}]
    else:
        k0 = f"{centers[0][0]},{centers[0][1]}"
        k1 = f"{centers[1][0]},{centers[1][1]}"
        a1, a2 = best_angles[k0][:2]
        b1, b2 = best_angles[k1][:2]
        combos = [
            {k0: a1, k1: b1},
            {k0: a2, k1: b1},
            {k0: a1, k1: b2},
            {k0: a2, k1: b2},
        ]

    images = []
    for combo in combos:
        tmp = svg_static
        for key, angle in combo.items():
            cx, cy = key.split(',')
            tmp = re.sub(rf'rotate\(\s*1\s*,\s*{re.escape(cx)}\s*,\s*{re.escape(cy)}\s*\)', f'rotate({angle}, {cx}, {cy})', tmp)
            tmp = re.sub(rf'rotate\(\s*-1\s*,\s*{re.escape(cx)}\s*,\s*{re.escape(cy)}\s*\)', f'rotate(-{angle}, {cx}, {cy})', tmp)
        try:
            png_bytes = svg2png(bytestring=tmp.encode('utf-8'))
            img = Image.open(io.BytesIO(png_bytes)).convert("RGBA")
            img = img.resize((600, 400))
            background = Image.new("RGBA", img.size, (255, 255, 255))
            combined = Image.alpha_composite(background, img).convert("RGB")
            images.append(combined)
        except Exception:
            continue

    ocr_results = []
    for img in images:
        try:
            txt = _single_ocr_from_image(img)
            ocr_results.append(txt)
        except Exception:
            ocr_results.append("")

    for r in ocr_results:
        if len(r) == 4:
            return r
    if ocr_results:
        best = max(ocr_results, key=lambda x: len(x or ""))
        return best or ""
    return ""

def predict(svgdata):
    if not svgdata:
        return "No SVG provided"
    if len(svgdata) > 50000:
        return "SVG too large"
    try:
        model_answer = solve_svg_captcha(svgdata)
    except Exception as e:
        print(f"Error in predict: {e}")
        return "Model could not predict"
    return model_answer or "Model could not predict"

with gr.Blocks() as demo:
    gr.Markdown("Enter SVG data and receive model answer")
    svg_input = gr.Textbox(label="SVG Data", lines=10)
    predict_btn = gr.Button("Get Model Answer")
    model_answer = gr.Textbox(label="Model Answer", interactive=False)
    predict_btn.click(predict, inputs=[svg_input], outputs=[model_answer])

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