app.py

import cv2
import mediapipe as mp
import numpy as np
import gradio as gr
import base64
import time
import os

# Set MPLCONFIGDIR to avoid Matplotlib cache permission issues
os.environ["MPLCONFIGDIR"] = "/tmp/matplotlib-cache"

# Initialize MediaPipe Selfie Segmentation
mp_selfie_segmentation = mp.solutions.selfie_segmentation
segmentation = mp_selfie_segmentation.SelfieSegmentation(model_selection=1)

# Global settings
settings = {
    "seg_enabled": True,
    "blur_bg": False,
    "set_bg": False,
    "set_color": False,
    "bg_color": (0, 0, 0),  # BGR
    "blur_intensity": 15
}
bg_image = None

def process_frame(frame, seg_enabled, blur_bg, set_bg, set_color, bg_color, blur_intensity, custom_image=None):
    global bg_image
    settings.update({
        "seg_enabled": seg_enabled,
        "blur_bg": blur_bg,
        "set_bg": set_bg,
        "set_color": set_color,
        "bg_color": tuple(map(int, bg_color.split(","))) if set_color else (0, 0, 0),
        "blur_intensity": blur_intensity
    })

    if custom_image is not None and set_bg:
        bg_image = custom_image

    if frame is None:
        return None, "No video frame received"

    process_start = time.time()
    frame_rgb = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
    result = segmentation.process(frame_rgb)
    mask = result.segmentation_mask

    # Create alpha mask
    alpha = mask > 0.5
    alpha = alpha.astype(np.uint8) * 255
    alpha = cv2.merge([alpha, alpha, alpha])

    output_frame = frame.copy()

    if settings["seg_enabled"]:
        if settings["blur_bg"]:
            bg = cv2.resize(frame, None, fx=0.1, fy=0.1, interpolation=cv2.INTER_LINEAR)
            ksize = settings["blur_intensity"]
            if ksize % 2 == 0:
                ksize -= 1
            bg = cv2.GaussianBlur(bg, (ksize, ksize), 0)
            bg = cv2.resize(bg, (frame.shape[1], frame.shape[0]), interpolation=cv2.INTER_LINEAR)
            output_frame = np.where(alpha == 255, frame, bg)
        elif settings["set_bg"] and bg_image is not None:
            if bg_image.shape[:2] != frame.shape[:2]:
                bg_image = cv2.resize(bg_image, (frame.shape[1], frame.shape[0]))
            output_frame = np.where(alpha == 255, frame, bg_image)
        elif settings["set_color"]:
            bg = np.full_like(frame, settings["bg_color"])
            output_frame = np.where(alpha == 255, frame, bg)
        else:
            bg = np.zeros_like(frame)
            output_frame = np.where(alpha == 255, frame, bg)

    process_time = (time.time() - process_start) * 1000
    return output_frame, f"{process_time:.2f} ms"

# Gradio interface
with gr.Blocks() as demo:
    gr.Markdown("# AI Background Remover")
    with gr.Row():
        with gr.Column():
            webcam = gr.Image(sources="webcam", label="Live Video", streaming=True)
            output_image = gr.Image(label="Processed Output")
            seg_enabled = gr.Checkbox(label="Enable Background Removal", value=True)
            blur_bg = gr.Checkbox(label="Blur Background")
            set_bg = gr.Checkbox(label="Custom Image Background")
            custom_image = gr.Image(label="Upload Custom Background")
            set_color = gr.Checkbox(label="Solid Color Background")
            bg_color = gr.Textbox(label="Background Color (R,G,B)", value="0,0,0")
            blur_intensity = gr.Slider(label="Blur Intensity", minimum=5, maximum=25, value=15, step=2)
            processing_time = gr.Textbox(label="Processing Time", value="0 ms")

    webcam.stream(
        fn=process_frame,
        inputs=[webcam, seg_enabled, blur_bg, set_bg, set_color, bg_color, blur_intensity, custom_image],
        outputs=[output_image, processing_time]
    )

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