app.py

import gradio as gr
from inference_sdk import InferenceHTTPClient
from PIL import Image, ImageDraw
import os
from collections import defaultdict

# ✅ Load API key securely from Hugging Face Spaces Secrets
API_KEY = os.getenv("ROBOFLOW_API_KEY")

if not API_KEY:
    raise ValueError("API Key is missing! Set it in HF Space Secrets.")

# ✅ Initialize Roboflow Client
CLIENT = InferenceHTTPClient(
    api_url="https://detect.roboflow.com",
    api_key=API_KEY
)

MODEL_ID = "hvacsym/5"
CONFIDENCE_THRESHOLD = 0.2  # ✅ Confidence threshold for filtering predictions
GRID_SIZE = (3, 3)  # ✅ 3x3 segmentation

def format_counts_as_table(counts, pass_num):
    """Formats detection counts into a Markdown table for Gradio."""
    if not counts:
        return f"### Pass {pass_num}: No components detected."
    
    table = f"### Pass {pass_num} Detection Results:\n\n"
    table += "| Component | Count |\n"
    table += "|-----------|-------|\n"
    for component, count in counts.items():
        table += f"| {component} | {count} |\n"
    
    return table

def detect_components(image):
    """ Detect components in an uploaded image with three passes. """
    
    original_image = image.convert("RGB")
    width, height = original_image.size
    seg_w, seg_h = width // GRID_SIZE[1], height // GRID_SIZE[0]
    
    def process_detection(image, pass_num):
        """ Detect objects in an image segment and remove them if found. """
        final_image = image.copy()
        draw_final = ImageDraw.Draw(final_image)
        total_counts = defaultdict(int)
        detected_boxes = []
        
        for row in range(GRID_SIZE[0]):
            for col in range(GRID_SIZE[1]):
                x1, y1 = col * seg_w, row * seg_h
                x2, y2 = (col + 1) * seg_w, (row + 1) * seg_h
                
                segment = image.crop((x1, y1, x2, y2))
                segment_path = f"segment_{row}_{col}_pass{pass_num}.png"
                segment.save(segment_path)
                
                # ✅ Run inference
                result = CLIENT.infer(segment_path, model_id=MODEL_ID)
                filtered_predictions = [pred for pred in result["predictions"] if pred["confidence"] >= CONFIDENCE_THRESHOLD]
                
                for obj in filtered_predictions:
                    sx, sy, sw, sh = obj["x"], obj["y"], obj["width"], obj["height"]
                    class_name = obj["class"]
                    total_counts[class_name] += 1

                    # ✅ Convert segment coordinates to full image coordinates
                    x_min_full, y_min_full = x1 + sx - sw // 2, y1 + sy - sh // 2
                    x_max_full, y_max_full = x1 + sx + sw // 2, y1 + sy + sh // 2
                    detected_boxes.append((x_min_full, y_min_full, x_max_full, y_max_full))
                    
                    # ✅ Draw bounding box
                    draw_final.rectangle([x_min_full, y_min_full, x_max_full, y_max_full], outline="green", width=2)

        return final_image, total_counts, detected_boxes

    # ✅ First pass detection
    image_after_pass1, counts_pass1, detected_boxes = process_detection(original_image, pass_num=1)
    counts_pass1_table = format_counts_as_table(counts_pass1, 1)

    # ✅ Mask detected areas for the second pass
    image_after_removal1 = original_image.copy()
    draw_removal1 = ImageDraw.Draw(image_after_removal1)
    for box in detected_boxes:
        draw_removal1.rectangle(box, fill=(255, 255, 255))

    # ✅ Second pass detection
    image_after_pass2, counts_pass2, detected_boxes = process_detection(image_after_removal1, pass_num=2)
    counts_pass2_table = format_counts_as_table(counts_pass2, 2)

    # ✅ Mask detected areas for the third pass
    image_after_removal2 = image_after_removal1.copy()
    draw_removal2 = ImageDraw.Draw(image_after_removal2)
    for box in detected_boxes:
        draw_removal2.rectangle(box, fill=(255, 255, 255))

    # ✅ Third pass detection
    image_after_pass3, counts_pass3, _ = process_detection(image_after_removal2, pass_num=3)
    counts_pass3_table = format_counts_as_table(counts_pass3, 3)

    # ✅ Sum counts from all passes
    final_counts = defaultdict(int)
    for label in set(counts_pass1) | set(counts_pass2) | set(counts_pass3):
        final_counts[label] = counts_pass1.get(label, 0) + counts_pass2.get(label, 0) + counts_pass3.get(label, 0)

    final_counts_table = format_counts_as_table(final_counts, "Final")

    # ✅ Return counts in Markdown table format
    return (
        image_after_pass1, counts_pass1_table,  
        image_after_pass2, counts_pass2_table,  
        image_after_pass3, counts_pass3_table,  
        final_counts_table
    )

# ✅ Gradio Interface
interface = gr.Interface(
    fn=detect_components,
    inputs=gr.Image(type="pil"),
    outputs=[
        gr.Image(type="pil", label="Detection Pass 1"),
        gr.Markdown(label="Counts After Pass 1"),  # ✅ Prettified Markdown Table
        gr.Image(type="pil", label="Detection Pass 2"),
        gr.Markdown(label="Counts After Pass 2"),  # ✅ Prettified Markdown Table
        gr.Image(type="pil", label="Detection Pass 3"),
        gr.Markdown(label="Counts After Pass 3"),  # ✅ Prettified Markdown Table
        gr.Markdown(label="Final Detected Components")  # ✅ Prettified Final Results
    ],
    title="HVAC Component Detector",
    description="Upload an image to detect HVAC components using Roboflow API across three passes."
)

# ✅ Launch the app
if __name__ == "__main__":
    interface.launch()