app.py

import gradio as gr
import numpy as np
import matplotlib.pyplot as plt
import matplotlib.colors as mcolors
from gradio_client import Client, handle_file
from PIL import Image
import requests
from io import BytesIO
import cv2

def get_segmentation_mask(image_url):
    client = Client("facebook/sapiens-seg")
    result = client.predict(image=handle_file(image_url), model_name="1b", api_name="/process_image")
    return np.load(result[1])  # Result[1] contains the .npy mask


def process_image(image, categories_to_hide):
    # Convert uploaded image to a PIL Image
    image = Image.open(image.name).convert("RGBA")
    
    # Save temporarily and get the segmentation mask
    image.save("temp_image.png")
    mask_data = get_segmentation_mask("temp_image.png")
    
    # Define grouped categories
    grouped_mapping = {
        "Background": [0],
        "Clothes": [1, 12, 22, 8, 9, 17, 18],  # Includes Shoes, Socks, Slippers
        "Face": [2, 23, 24, 25, 26, 27],  # Face, Neck, Lips, Teeth, Tongue
        "Hair": [3],  # Hair
        "Skin": [4, 5, 6, 7, 10, 11, 13, 14, 15, 16, 19, 20, 21]  # Hands, Feet, Arms, Legs, Torso
    }

    # Convert image to numpy array (RGBA)
    image_array = np.array(image, dtype=np.uint8)
    
    # Create an empty transparent image
    transparent_image = np.zeros_like(image_array, dtype=np.uint8)
    
    # Create a binary mask for selected categories
    mask_combined = np.zeros_like(mask_data, dtype=bool)

    for category in categories_to_hide:
        for idx in grouped_mapping.get(category, []):
            mask_combined |= (mask_data == idx)

    # Expand clothing boundaries if clothes are in `categories_to_hide`
    if "Clothes" in categories_to_hide:
        clothing_mask = np.isin(mask_data, grouped_mapping["Clothes"]).astype(np.uint8)
    
        # Determine kernel size (2% of the smaller image dimension)
        height, width = clothing_mask.shape
        kernel_size = max(20, int(0.02 * min(height, width)))  # Ensure at least 1 pixel
        kernel = np.ones((kernel_size, kernel_size), np.uint8)
    
        # **Step 1: Fill gaps using Morphological Closing (Dilation + Erosion)**
        closed_clothing_mask = cv2.morphologyEx(clothing_mask, cv2.MORPH_CLOSE, kernel, iterations=1)
    
        # **Step 2: Expand clothing boundary using Dilation**
        dilated_clothing_mask = cv2.dilate(closed_clothing_mask, kernel, iterations=1)
    
        # Update mask_combined with the expanded clothing mask
        mask_combined |= (dilated_clothing_mask == 1)


    # Apply the mask (preserve only selected regions)
    transparent_image[mask_combined] = image_array[mask_combined]
    
    # Convert back to PIL Image
    result_image = Image.fromarray(transparent_image, mode="RGBA")
    
    return result_image

# Define Gradio Interface
demo = gr.Interface(
    fn=process_image,
    inputs=[
        gr.File(label="Upload an Image"),
        gr.CheckboxGroup([
            "Background", "Clothes", "Face", "Hair", "Skin"
        ], label="Select Categories to Preserve")
    ],
    outputs=gr.Image(label="Masked Image", type="pil"),
    title="Segmentation Mask Editor",
    description="Upload an image, generate a segmentation mask, and select categories to preserve while making the rest transparent."
)

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