app.py

import os
import sys

# Standard environment setup (keep this)
if "APP_PATH" in os.environ:
    app_path = os.path.abspath(os.environ["APP_PATH"])
    if os.getcwd() != app_path:
        # fix sys.path for import
        os.chdir(app_path)
    if app_path not in sys.path:
        sys.path.append(app_path)

import io
import tempfile
from typing import List

import pypdfium2
import gradio as gr
import requests
from contextlib import suppress

from surya.common.surya.schema import TaskNames
from surya.models import load_predictors

from surya.debug.draw import draw_polys_on_image
from PIL import Image
from surya.layout import LayoutResult
from surya.settings import settings
from surya.common.util import rescale_bbox, expand_bbox


# --- Core Functions (Minimal changes required) ---

# Get page image from PDF (keep this)
def open_pdf(pdf_file):
    return pypdfium2.PdfDocument(pdf_file)

def page_counter(pdf_file):
    doc = open_pdf(pdf_file)
    doc_len = len(doc)
    doc.close()
    return doc_len

def get_page_image(pdf_file, page_num, dpi=settings.IMAGE_DPI):
    doc = open_pdf(pdf_file)
    renderer = doc.render(
        pypdfium2.PdfBitmap.to_pil,
        page_indices=[page_num - 1],
        scale=dpi / 72,
    )
    png = list(renderer)[0]
    png_image = png.convert("RGB")
    doc.close()
    return png_image

def get_uploaded_image(in_file):
    return Image.open(in_file).convert("RGB")

# Modified layout_detection to filter for Equation and Figure
def focused_layout_detection(img) -> (Image.Image, LayoutResult):
    # Use the existing layout predictor
    pred = predictors["layout"]([img])[0]
    
    # Filter for Equation and Figure bounding boxes
    filtered_bboxes = [
        p
        for p in pred.bboxes
        if p.label in ["Equation", "Figure"] # <-- Filter applied here
    ]
    
    # Update the prediction result to only include the filtered boxes
    pred.bboxes = filtered_bboxes
    
    # Prepare data for drawing on the image
    polygons = [p.polygon for p in filtered_bboxes]
    labels = [
        f"{p.label}-{p.position}-{round(p.top_k[p.label], 2)}" for p in filtered_bboxes
    ]
    
    # Draw the filtered polygons
    layout_img = draw_polys_on_image(
        polygons, img.copy(), labels=labels, label_font_size=18
    )
    
    return layout_img, pred


# Load models (keep this)
predictors = load_predictors()


# --- Gradio Interface (Significantly simplified) ---

with gr.Blocks(title="Surya Equation/Figure Detector") as demo:
    gr.Markdown("""
    # Surya Equation and Figure Detection

    This application uses Surya OCR's layout analysis model to **specifically detect and locate Equations and Figures** within a document page.

    The output provides an image with bounding boxes drawn, and the raw JSON bounding box information for the detected elements.

    Find the original project [here](https://github.com/VikParuchuri/surya).
    """)

    with gr.Row():
        with gr.Column():
            in_file = gr.File(label="PDF file or image:", file_types=[".pdf", ".png", ".jpg", ".jpeg", ".gif", ".webp"])
            in_num = gr.Slider(label="Page number", minimum=1, maximum=100, value=1, step=1)
            in_img = gr.Image(label="Select page of Image", type="pil", sources=None)

            # Keep only the essential button
            detection_btn = gr.Button("Run Equation and Figure Detection")
            
        with gr.Column():
            result_img = gr.Gallery(label="Result image: Detected Equations and Figures", show_label=True, 
                elem_id="gallery", columns=[1], rows=[1], object_fit="contain", height="auto")

            gr.HTML("""
            <style>
            #gallery {
                height: auto !important;
                max-height: none !important;
                overflow: visible !important;
            }
            #gallery .gallery-item {
                flex-direction: column !important;
            }
            #gallery .gallery-item img {
                width: 100% !important;
                height: auto !important;
                object-fit: contain !important;
            }
            </style>
            """)
            result_json = gr.JSON(label="Result JSON (Bounding Box Data)")

        # Page Loading Logic (keep this)
        def show_image(file, num=1):
            if file.endswith('.pdf'):
                count = page_counter(file)
                img = get_page_image(file, num, settings.IMAGE_DPI)
                return [
                    gr.update(visible=True, maximum=count),
                    gr.update(value=img)]
            else:
                img = get_uploaded_image(file)
                return [
                    gr.update(visible=False),
                    gr.update(value=img)]

        in_file.upload(
            fn=show_image,
            inputs=[in_file],
            outputs=[in_num, in_img],
        )
        in_num.change(
            fn=show_image,
            inputs=[in_file, in_num],
            outputs=[in_num, in_img],
        )

        # Run Focused Detection
        def run_focused_detection(pil_image):
            # update counter
            with suppress(Exception):
                requests.get("https://counterapi.com/api/xiaoyao9184.github.com/view/docker-surya")

            layout_img, pred = focused_layout_detection(pil_image)
            # Exclude the large segmentation map from the JSON output
            layout_json = pred.model_dump(exclude=["segmentation_map"]) 
            
            # Count the filtered results
            num_boxes = len(layout_json.get('bboxes', []))
            
            return (
                gr.update(label=f"Result image: {num_boxes} Equations/Figures detected", value=[layout_img], rows=[1], height=layout_img.height),
                gr.update(label=f"Result JSON: {num_boxes} Equations/Figures detected", value=layout_json)
            )
        
        detection_btn.click(
            fn=run_focused_detection,
            inputs=[in_img],
            outputs=[result_img, result_json]
        )


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