MogensR's picture
Update processing/matting.py
30d0dd7
raw
history blame
13.5 kB
#!/usr/bin/env python3
"""
Advanced matting algorithms for BackgroundFX Pro.
Implements multiple matting techniques with automatic fallback.
"""
from dataclasses import dataclass
from typing import Dict, Optional
import cv2
import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
from utils.logger import get_logger
from utils.hardware.device_manager import DeviceManager
from utils.config import ConfigManager # kept for forward compatibility / config hook
from core.models import ModelFactory, ModelType # not used directly here but kept for API consistency
from core.quality import QualityAnalyzer
from core.edge import EdgeRefinement
logger = get_logger(__name__)
@dataclass
class MattingConfig:
"""Configuration for matting operations."""
alpha_threshold: float = 0.5
erode_iterations: int = 2
dilate_iterations: int = 2
blur_radius: int = 3
trimap_size: int = 30
confidence_threshold: float = 0.7
use_guided_filter: bool = True
guided_filter_radius: int = 8
guided_filter_eps: float = 1e-6
use_temporal_smoothing: bool = False
temporal_window: int = 5
class AlphaMatting:
"""Advanced alpha matting using multiple techniques."""
def __init__(self, config: Optional[MattingConfig] = None):
self.config = config or MattingConfig()
self.device_manager = DeviceManager()
self.quality_analyzer = QualityAnalyzer()
self.edge_refinement = EdgeRefinement()
def create_trimap(self, mask: np.ndarray, dilation_size: Optional[int] = None) -> np.ndarray:
"""
Create trimap from a binary mask.
Args:
mask: Binary mask (H, W) in {0, 255} or [0,1]
dilation_size: Size of uncertain region (pixels)
Returns:
Trimap with values 0 (background), 128 (unknown), 255 (foreground)
"""
dilation_size = dilation_size or self.config.trimap_size
# Ensure uint8 binary
if mask.dtype != np.uint8:
mask = (mask * 255).astype(np.uint8)
mask = (mask > 127).astype(np.uint8) * 255
trimap = np.copy(mask)
kernel = cv2.getStructuringElement(cv2.MORPH_ELLIPSE, (dilation_size, dilation_size))
# Dilate/erode once to form unknown band
dilated = cv2.dilate(mask, kernel, iterations=1)
eroded = cv2.erode(mask, kernel, iterations=1)
# Unknown where dilation has expanded FG beyond eroded FG band
trimap[:] = 0
trimap[eroded == 255] = 255
unknown = (dilated == 255) & (eroded == 0)
trimap[unknown] = 128
return trimap
def guided_filter(
self,
image: np.ndarray,
guide: np.ndarray,
radius: Optional[int] = None,
eps: Optional[float] = None,
) -> np.ndarray:
"""
Apply guided filter for edge-preserving smoothing.
Args:
image: Input image to filter (H, W) uint8
guide: Guide image (H, W, 3) or (H, W)
radius: Filter radius
eps: Regularization parameter
Returns:
Filtered image (H, W) uint8
"""
radius = radius or self.config.guided_filter_radius
eps = eps or self.config.guided_filter_eps
if guide.ndim == 3:
guide_gray = cv2.cvtColor(guide, cv2.COLOR_BGR2GRAY)
else:
guide_gray = guide
# Convert to float32 in [0,1]
I = guide_gray.astype(np.float32) / 255.0
p = image.astype(np.float32) / 255.0
# Box filter helper
def box_filter(img, r):
return cv2.boxFilter(img, -1, (r, r))
mean_I = box_filter(I, radius)
mean_p = box_filter(p, radius)
mean_Ip = box_filter(I * p, radius)
cov_Ip = mean_Ip - mean_I * mean_p
mean_II = box_filter(I * I, radius)
var_I = mean_II - mean_I * mean_I
a = cov_Ip / (var_I + eps)
b = mean_p - a * mean_I
mean_a = box_filter(a, radius)
mean_b = box_filter(b, radius)
q = mean_a * I + mean_b
return np.clip(q * 255.0, 0, 255).astype(np.uint8)
def closed_form_matting(self, image: np.ndarray, trimap: np.ndarray) -> np.ndarray:
"""
Closed-form-inspired fast matting using distance transforms + optional guided filtering.
Args:
image: RGB image (H, W, 3) uint8
trimap: Trimap with values {0, 128, 255}
Returns:
Alpha matte in [0,1] float32
"""
h, w = trimap.shape[:2]
alpha = (trimap.astype(np.float32) / 255.0)
is_fg = trimap == 255
is_bg = trimap == 0
is_unknown = trimap == 128
if not np.any(is_unknown):
return np.clip(alpha, 0.0, 1.0)
dist_fg = cv2.distanceTransform(is_fg.astype(np.uint8), cv2.DIST_L2, 5)
dist_bg = cv2.distanceTransform(is_bg.astype(np.uint8), cv2.DIST_L2, 5)
total = dist_fg + dist_bg + 1e-10
alpha_unknown = dist_fg / total
alpha[is_unknown] = alpha_unknown[is_unknown]
if self.config.use_guided_filter:
alpha_u8 = np.clip(alpha * 255.0, 0, 255).astype(np.uint8)
alpha_u8 = self.guided_filter(alpha_u8, image)
alpha = alpha_u8.astype(np.float32) / 255.0
return np.clip(alpha, 0.0, 1.0)
def deep_matting(
self,
image: np.ndarray,
mask: np.ndarray,
model: Optional[nn.Module] = None,
) -> np.ndarray:
"""
Apply deep learning-based matting refinement.
Args:
image: RGB image (H, W, 3) uint8
mask: Initial mask (H, W) {0..255} or [0,1]
model: Optional pre-trained model taking (img, mask) → alpha
Returns:
Refined alpha matte in [0,1] float32
"""
device = self.device_manager.get_device()
h, w = image.shape[:2]
input_size = (512, 512)
img_rs = cv2.resize(image, input_size)
msk_rs = cv2.resize(mask, input_size)
img_t = torch.from_numpy(img_rs.transpose(2, 0, 1)).float().unsqueeze(0) / 255.0
msk_t = torch.from_numpy(msk_rs).float().unsqueeze(0).unsqueeze(0)
if msk_t.max() > 1.0:
msk_t = msk_t / 255.0
img_t = img_t.to(device)
msk_t = msk_t.to(device)
with torch.no_grad():
if model is None:
x = torch.cat([img_t, msk_t], dim=1)
refined = self._simple_refine_network(x)
else:
refined = model(img_t, msk_t)
alpha = refined.squeeze().float().cpu().numpy()
alpha = cv2.resize(alpha, (w, h))
return np.clip(alpha, 0.0, 1.0)
def _simple_refine_network(self, x: torch.Tensor) -> torch.Tensor:
"""Tiny non-learned refinement block (demo-quality)."""
# x: [B, 4, H, W] (RGB + mask)
mask = x[:, 3:4, :, :]
refined = mask
for _ in range(3):
refined = F.avg_pool2d(refined, 3, stride=1, padding=1)
refined = torch.sigmoid((refined - 0.5) * 10.0)
return refined
def morphological_refinement(self, alpha: np.ndarray) -> np.ndarray:
"""
Apply morphological operations and boundary smoothing.
Args:
alpha: Alpha matte in [0,1] float32
Returns:
Refined alpha in [0,1] float32
"""
alpha_u8 = np.clip(alpha * 255.0, 0, 255).astype(np.uint8)
kernel = cv2.getStructuringElement(cv2.MORPH_ELLIPSE, (5, 5))
# Close small holes in FG
alpha_u8 = cv2.morphologyEx(
alpha_u8, cv2.MORPH_CLOSE, kernel, iterations=self.config.erode_iterations
)
# Remove small specks
alpha_u8 = cv2.morphologyEx(
alpha_u8, cv2.MORPH_OPEN, kernel, iterations=self.config.dilate_iterations
)
if self.config.blur_radius > 0:
r = self.config.blur_radius * 2 + 1
alpha_u8 = cv2.GaussianBlur(alpha_u8, (r, r), 0)
return alpha_u8.astype(np.float32) / 255.0
def process(self, image: np.ndarray, mask: np.ndarray, method: str = "auto") -> Dict[str, np.ndarray]:
"""
Process image with selected matting method.
Args:
image: RGB image (H, W, 3) uint8
mask: Initial segmentation mask (H, W)
method: 'auto' | 'trimap' | 'deep' | 'guided'
Returns:
dict(alpha, confidence, method_used, quality_metrics[, error])
"""
try:
quality_metrics = self.quality_analyzer.analyze_frame(image)
chosen = method
if method == "auto":
# Heuristic selection
blur_score = quality_metrics.get("blur_score", 0.0)
edge_clarity = quality_metrics.get("edge_clarity", 0.0)
if blur_score > 50:
chosen = "guided"
elif edge_clarity > 0.7:
chosen = "trimap"
else:
chosen = "deep"
logger.info(f"Using matting method: {chosen}")
if chosen == "trimap":
trimap = self.create_trimap(mask)
alpha = self.closed_form_matting(image, trimap)
elif chosen == "deep":
alpha = self.deep_matting(image, mask)
elif chosen == "guided":
alpha = mask.astype(np.float32)
if alpha.max() > 1.0:
alpha = alpha / 255.0
if self.config.use_guided_filter:
alpha_u8 = np.clip(alpha * 255.0, 0, 255).astype(np.uint8)
alpha = self.guided_filter(alpha_u8, image).astype(np.float32) / 255.0
else:
alpha = mask.astype(np.float32)
if alpha.max() > 1.0:
alpha = alpha / 255.0
# Morphological + edge refinement
alpha = self.morphological_refinement(alpha)
alpha = self.edge_refinement.refine_edges(
image, np.clip(alpha * 255.0, 0, 255).astype(np.uint8)
).astype(np.float32) / 255.0
confidence = self._calculate_confidence(alpha, quality_metrics)
return {
"alpha": np.clip(alpha, 0.0, 1.0),
"confidence": float(np.clip(confidence, 0.0, 1.0)),
"method_used": chosen,
"quality_metrics": quality_metrics,
}
except Exception as e:
logger.error(f"Matting processing failed: {e}")
fallback = mask.astype(np.float32)
if fallback.max() > 1.0:
fallback = fallback / 255.0
return {
"alpha": np.clip(fallback, 0.0, 1.0),
"confidence": 0.0,
"method_used": "fallback",
"error": str(e),
}
def _calculate_confidence(self, alpha: np.ndarray, quality_metrics: Dict) -> float:
"""Calculate confidence score for the matting result."""
confidence = float(quality_metrics.get("overall_quality", 0.5))
alpha_mean = float(np.mean(alpha))
alpha_std = float(np.std(alpha))
# Prefer clear separation
if 0.3 < alpha_mean < 0.7 and alpha_std > 0.3:
confidence *= 1.2
edges = cv2.Canny(np.clip(alpha * 255.0, 0, 255).astype(np.uint8), 50, 150)
edge_ratio = float(np.sum(edges > 0) / edges.size)
if edge_ratio < 0.1:
confidence *= 1.1
return float(np.clip(confidence, 0.0, 1.0))
class CompositingEngine:
"""Handle alpha compositing and blending."""
def __init__(self):
self.logger = get_logger(f"{__name__}.CompositingEngine")
def composite(self, foreground: np.ndarray, background: np.ndarray, alpha: np.ndarray) -> np.ndarray:
"""
Composite foreground over background using alpha.
Args:
foreground: Foreground image (H, W, 3) uint8
background: Background image (H, W, 3) uint8
alpha: Alpha matte (H, W) or (H, W, 1) in [0..255] or [0..1]
Returns:
Composited image (H, W, 3) uint8
"""
# Ensure alpha is 3-channel
if alpha.ndim == 2:
alpha = np.expand_dims(alpha, axis=2)
if alpha.shape[2] == 1:
alpha = np.repeat(alpha, 3, axis=2)
# Normalize alpha to [0,1]
a = alpha.astype(np.float32)
if a.max() > 1.0:
a = a / 255.0
fg = foreground.astype(np.float32) / 255.0
bg = background.astype(np.float32) / 255.0
result = fg * a + bg * (1.0 - a)
return np.clip(result * 255.0, 0, 255).astype(np.uint8)
def premultiply_alpha(self, image: np.ndarray, alpha: np.ndarray) -> np.ndarray:
"""
Premultiply RGB image by alpha channel.
Args:
image: (H, W, 3) uint8
alpha: (H, W) or (H, W, 1) in [0..255] or [0..1]
Returns:
Premultiplied (H, W, 3) uint8
"""
if alpha.ndim == 2:
alpha = np.expand_dims(alpha, axis=2)
if alpha.shape[2] == 1:
alpha = np.repeat(alpha, 3, axis=2)
a = alpha.astype(np.float32)
if a.max() > 1.0:
a = a / 255.0
img_f = image.astype(np.float32)
premul = img_f * a
return np.clip(premul, 0.0, 255.0).astype(np.uint8)