SegBody.py

from transformers import pipeline
import numpy as np
import cv2
from PIL import Image, ImageDraw

# Try to import insightface for face detection, fallback to OpenCV if not available
try:
    from insightface.app import FaceAnalysis
    USE_INSIGHTFACE = True
except ImportError:
    USE_INSIGHTFACE = False
    print("⚠️ InsightFace not available, using OpenCV for face detection")

# Initialize face detection
if USE_INSIGHTFACE:
    try:
        app = FaceAnalysis(providers=['CPUExecutionProvider'])  # CPU only for HF Spaces
        app.prepare(ctx_id=-1, det_size=(640, 640))  # -1 for CPU
    except Exception as e:
        print(f"⚠️ InsightFace initialization failed: {e}, falling back to OpenCV")
        USE_INSIGHTFACE = False

# Initialize segmentation pipeline
segmenter = pipeline(model="mattmdjaga/segformer_b2_clothes", device=-1)  # -1 for CPU


def remove_face_opencv(img, mask):
    """Fallback face detection using OpenCV"""
    img_arr = np.asarray(img.convert('RGB'))
    gray = cv2.cvtColor(img_arr, cv2.COLOR_RGB2GRAY)
    
    # Load OpenCV's pre-trained face detector
    face_cascade = cv2.CascadeClassifier(cv2.data.haarcascades + 'haarcascade_frontalface_default.xml')
    
    # Detect faces
    faces = face_cascade.detectMultiScale(gray, 1.1, 4)
    
    if len(faces) > 0:
        # Get the first (largest) face
        x, y, w, h = faces[0]
        
        # Expand face region
        x = max(0, int(x - w * 0.5))
        y = max(0, int(y - h * 0.5))
        w = int(w * 2.0)
        h = int(h * 1.7)
        
        # Draw black rectangle on mask
        img1 = ImageDraw.Draw(mask)
        img1.rectangle([(x, y), (x + w, y + h)], fill=0)
    
    return mask


def remove_face(img, mask):
    """Remove face from mask using InsightFace or OpenCV fallback"""
    if not USE_INSIGHTFACE:
        return remove_face_opencv(img, mask)
    
    try:
        # Convert image to numpy array
        img_arr = np.asarray(img)
        
        # Run face detection
        faces = app.get(img_arr)
        
        if len(faces) == 0:
            return mask
        
        # Get the first face
        bbox = faces[0]['bbox']

        # Width and height of face
        w = bbox[2] - bbox[0]
        h = bbox[3] - bbox[1]

        # Make face locations bigger
        bbox[0] = bbox[0] - (w * 0.5)  # x left
        bbox[2] = bbox[2] + (w * 0.5)  # x right
        bbox[1] = bbox[1] - (h * 0.5)  # y top
        bbox[3] = bbox[3] + (h * 0.2)  # y bottom

        # Convert to [(x_left, y_top), (x_right, y_bottom)]
        face_locations = [(bbox[0], bbox[1]), (bbox[2], bbox[3])]

        # Draw black rect onto mask
        img1 = ImageDraw.Draw(mask)
        img1.rectangle(face_locations, fill=0)

        return mask
    except Exception as e:
        print(f"⚠️ InsightFace face removal failed: {e}, using OpenCV")
        return remove_face_opencv(img, mask)


def segment_body(original_img, face=True):
    """
    Segment body from image
    
    Args:
        original_img: PIL Image
        face: If True, includes face in mask. If False, excludes face.
    
    Returns:
        tuple: (segmented_image, mask_image)
    """
    # Make a copy
    img = original_img.copy()
    
    # Segment image
    segments = segmenter(img)

    # Create list of masks
    segment_include = ["Hat", "Hair", "Sunglasses", "Upper-clothes", "Skirt", "Pants", 
                      "Dress", "Belt", "Left-shoe", "Right-shoe", "Face", "Left-leg", 
                      "Right-leg", "Left-arm", "Right-arm", "Bag", "Scarf"]
    mask_list = []
    for s in segments:
        if s['label'] in segment_include:
            mask_list.append(s['mask'])

    if len(mask_list) == 0:
        # If no segments found, return full mask
        print("⚠️ No body segments detected, using full mask")
        final_mask = Image.new('L', img.size, 255)
    else:
        # Paste all masks on top of each other 
        final_mask = np.array(mask_list[0])
        for mask in mask_list[1:]:
            current_mask = np.array(mask)
            final_mask = final_mask + current_mask
                
        # Convert final mask from np array to PIL image
        final_mask = Image.fromarray(final_mask)

    # Remove face
    if face == False:
        final_mask = remove_face(img.convert('RGB'), final_mask)

    # Apply mask to original image
    img_copy = img.copy()
    img_copy.putalpha(final_mask)

    return img_copy, final_mask


def segment_torso(original_img):
    """
    Segment only torso/upper body from image
    
    Args:
        original_img: PIL Image
    
    Returns:
        tuple: (segmented_image, mask_image)
    """
    # Make a copy
    img = original_img.copy()
    
    # Segment image
    segments = segmenter(img)

    # Create list of masks (torso only)
    segment_include = ["Upper-clothes", "Dress", "Belt", "Face", "Left-arm", "Right-arm"]
    mask_list = []
    for s in segments:
        if s['label'] in segment_include:
            mask_list.append(s['mask'])

    if len(mask_list) == 0:
        # If no segments found, return full mask
        print("⚠️ No torso segments detected, using full mask")
        final_mask = Image.new('L', img.size, 255)
    else:
        # Paste all masks on top of each other 
        final_mask = np.array(mask_list[0])
        for mask in mask_list[1:]:
            current_mask = np.array(mask)
            final_mask = final_mask + current_mask
                
        # Convert final mask from np array to PIL image
        final_mask = Image.fromarray(final_mask)

    # Remove face
    final_mask = remove_face(img.convert('RGB'), final_mask)

    # Apply mask to original image
    img_copy = img.copy()
    img_copy.putalpha(final_mask)

    return img_copy, final_mask