vichetkao's picture
Update app.py
a4ed4be verified
Raw
History Blame Contribute Delete
10.2 kB
import os
import io
import json
from pathlib import Path
import requests
import numpy as np
import gradio as gr
from PIL import Image, ImageDraw, ImageFont
from pipeline import create_labelme_json, clean_labelme
# ---------------------------------------------------------------------------
# Config
# ---------------------------------------------------------------------------
# Hosted Ultralytics inference endpoint. Prefer setting these as Space secrets
# (env vars); the values below are fallbacks so it runs out of the box.
API_URL = os.getenv("API_URL")
API_KEY = os.getenv("API_KEY")
IMAGE_FOLDER = "images"
# Color per class keyword (RGB)
CLASS_COLORS = {
'column': (255, 165, 0), # orange
'row': (0, 200, 0), # green
'header': (30, 120, 255), # blue
'line': (230, 230, 0), # yellow
}
DEFAULT_COLOR = (255, 0, 0) # red
def color_for_label(label):
low = label.lower()
for key, color in CLASS_COLORS.items():
if key in low:
return color
return DEFAULT_COLOR
# ---------------------------------------------------------------------------
# Test images
# ---------------------------------------------------------------------------
def get_test_images():
images = []
if os.path.exists(IMAGE_FOLDER):
for file in sorted(Path(IMAGE_FOLDER).glob("*")):
if file.suffix.lower() in [".jpg", ".jpeg", ".png", ".bmp", ".gif"]:
images.append((str(file), file.name))
return images
def load_test_image(image_path):
if image_path and os.path.exists(image_path):
return Image.open(image_path).convert("RGB")
return None
# ---------------------------------------------------------------------------
# Drawing
# ---------------------------------------------------------------------------
def _load_font(font_size):
font_paths = [
"/usr/share/fonts/truetype/dejavu/DejaVuSans-Bold.ttf",
"/usr/share/fonts/truetype/liberation/LiberationSans-Bold.ttf",
"/System/Library/Fonts/Arial.ttf",
"C:\\Windows\\Fonts\\arial.ttf",
"arial.ttf",
]
for path in font_paths:
if os.path.exists(path):
try:
return ImageFont.truetype(path, font_size)
except Exception:
continue
return ImageFont.load_default()
def draw_shapes_on_image(image, shapes):
"""Draw cleaned labelme rectangle shapes onto a PIL image."""
if not shapes:
return image
img = image.copy()
draw = ImageDraw.Draw(img)
img_w, img_h = img.size
min_dim = min(img_w, img_h)
font_size = max(int(min_dim * 0.018), 16)
line_width = max(int(min_dim * 0.004), 2)
font = _load_font(font_size)
for shape in shapes:
a = np.array(shape["points"])
x1, y1 = int(np.min(a[:, 0])), int(np.min(a[:, 1]))
x2, y2 = int(np.max(a[:, 0])), int(np.max(a[:, 1]))
label = shape["label"]
color = color_for_label(label)
if x2 <= x1 or y2 <= y1:
continue
draw.rectangle([x1, y1, x2, y2], outline=color, width=line_width)
bbox = draw.textbbox((0, 0), label, font=font)
tw, th = bbox[2] - bbox[0], bbox[3] - bbox[1]
ty = max(0, y1 - th - 6)
pad = 3
draw.rectangle([x1, ty, x1 + tw + 2 * pad, ty + th + 2 * pad], fill=(0, 0, 0))
draw.text((x1 + pad, ty + pad), label, font=font, fill=color)
return img
# ---------------------------------------------------------------------------
# Prediction
# ---------------------------------------------------------------------------
def format_results(shapes, img_w, img_h):
out = "## Detection Results\n\n"
out += f"**Image Size:** {img_w} x {img_h} (W x H)\n\n"
out += f"**Shapes Found:** {len(shapes)}\n\n"
if shapes:
out += "### Detected Objects\n"
out += "| Label | Confidence |\n"
out += "|-------|------------|\n"
for s in shapes:
desc = s.get("description", "")
conf = desc.replace("confidence:", "").strip() if desc else "N/A"
out += f"| {s['label']} | {conf} |\n"
return out
def call_api(image, confidence, iou, imgsz):
"""POST the image to the hosted Ultralytics endpoint and return the JSON."""
img_bytes = io.BytesIO()
image.save(img_bytes, format="JPEG")
img_bytes.seek(0)
params = {"conf": confidence, "iou": iou, "imgsz": imgsz}
headers = {"Authorization": f"Bearer {API_KEY}"}
files = {"file": ("image.jpg", img_bytes, "image/jpeg")}
response = requests.post(API_URL, headers=headers, data=params, files=files, timeout=60)
response.raise_for_status()
return response.json()
def api_results_to_detections(api_result):
"""Convert the API response into the pipeline's detections dict."""
boxes = []
images = api_result.get("images", []) if isinstance(api_result, dict) else []
if images:
for det in images[0].get("results", []):
box = det.get("box", {})
x1 = float(box.get("x1", 0))
y1 = float(box.get("y1", 0))
x2 = float(box.get("x2", 0))
y2 = float(box.get("y2", 0))
boxes.append({
"points": [[x1, y1], [x2, y1], [x2, y2], [x1, y2]],
"confidence": float(det.get("confidence", 0)),
"class_name": det.get("name", "unknown"),
"class_id": int(det.get("class", 0)),
})
return {"boxes": boxes}
def predict_image(image, confidence, iou, imgsz):
if image is None:
return None, None, "#### Please upload an image to begin detection"
try:
image = image.convert("RGB")
api_result = call_api(image, float(confidence), float(iou), int(imgsz))
detections = api_results_to_detections(api_result)
# Build + clean labelme JSON (rows span columns, columns span header->last row, dedupe)
labelme_json = create_labelme_json(
"image.png", detections, image.height, image.width)
labelme_json = clean_labelme(labelme_json)
shapes = labelme_json["shapes"]
result_img = draw_shapes_on_image(image, shapes)
report = format_results(shapes, image.width, image.height)
json_path = os.path.join(os.getcwd(), "result.json")
with open(json_path, "w", encoding="utf-8") as f:
json.dump(labelme_json, f, indent=2)
return result_img, json_path, report
except requests.exceptions.Timeout:
return None, None, "#### Error: Request timeout. Please try again."
except requests.exceptions.ConnectionError:
return None, None, "#### Error: Unable to connect to detection service."
except requests.exceptions.HTTPError as e:
return None, None, f"#### Error: API returned status {e.response.status_code}"
except Exception as e:
return None, None, f"#### Error: {str(e)}"
# ---------------------------------------------------------------------------
# UI
# ---------------------------------------------------------------------------
dark_theme = gr.themes.Monochrome(
primary_hue="slate",
secondary_hue="slate",
).set(
body_text_color="#e0e0e0",
background_fill_primary="#0f0f0f",
background_fill_secondary="#1a1a1a",
)
with gr.Blocks(title="Table Layout Detection") as demo:
gr.Markdown("""
# Table Layout Detection
Detect table columns, rows and headers. Upload an image and adjust the
inference parameters. Boxes are auto-cleaned (rows span all columns, columns
span header→last row, duplicates removed) before being drawn.
""")
with gr.Row():
with gr.Column(scale=1, min_width=400):
gr.Markdown("### Input")
image_input = gr.Image(label="Image", type="pil", sources=["upload"], interactive=True)
test_images = get_test_images()
if test_images:
test_image_radio = gr.Radio(
choices=[img[1] for img in test_images],
label="Select test image", info="Click to load",
)
test_image_radio.change(
fn=lambda name: load_test_image(
next((img[0] for img in test_images if img[1] == name), None)),
inputs=[test_image_radio], outputs=[image_input],
)
else:
gr.Markdown("No test images found. Add images to the 'images' folder.")
gr.Markdown("### Configuration")
confidence_slider = gr.Slider(label="Confidence Threshold", minimum=0.0,
maximum=1.0, value=0.2, step=0.01,
info="Detection confidence level")
iou_slider = gr.Slider(label="IOU Threshold (NMS)", minimum=0.0, maximum=1.0,
value=0.2, step=0.01,
info="Intersection over union threshold")
imgsz_slider = gr.Slider(label="Image Size", minimum=320, maximum=2048,
value=1280, step=32, info="Inference image resolution")
predict_btn = gr.Button("Detect Objects", variant="primary", size="lg")
with gr.Column(scale=1, min_width=400):
gr.Markdown("### Results")
image_output = gr.Image(label="Detections", type="pil", interactive=False)
json_output = gr.File(label="Download labelme JSON")
results_output = gr.Markdown(value="Detection results will appear here.")
inputs = [image_input, confidence_slider, iou_slider, imgsz_slider]
outputs = [image_output, json_output, results_output]
predict_btn.click(fn=predict_image, inputs=inputs, outputs=outputs)
image_input.change(fn=predict_image, inputs=inputs, outputs=outputs)
confidence_slider.change(fn=predict_image, inputs=inputs, outputs=outputs)
iou_slider.change(fn=predict_image, inputs=inputs, outputs=outputs)
imgsz_slider.change(fn=predict_image, inputs=inputs, outputs=outputs)
if __name__ == "__main__":
demo.launch(share=False, show_error=True,
theme=dark_theme, css="footer {display: none !important;}")