Files
BGfilter/bgfilter/trimap.py
T
lhk229 adbbb9d620 Add cross-model veto of hair-gap background residue (default on)
Colour-drifted background trapped between hair strands defeats every
single-signal defence: the chroma key reads it as foreground (bgc ~0.07),
the segmenter backs it, ViTMatte rates it opaque, and post-hoc removal is
a proven dead end (it shreds the hair volume the same pixels belong to).
A second, trimap-free matting model (BiRefNet_HR-matting) is the only
tested model that separates this residue from the subject, so its opinion
is fused in as a veto: min-fusion that may only LOWER alpha, restricted to
the background-hued bright suspect zone (proj >= 3, L >= 45, feathered)
and gated by primary-alpha confidence (0.70 -> 0.95 ramp) so soft wisps
and dark hair are exempt by construction.

- settings/config/CLI: cross_check block, --cross-check/--no-cross-check
- alpha_post.cross_check_alpha after clean_alpha; second opinion reuses
  BiRefNetSegmenter; saved to debug as cross_check_alpha.png
- chroma.bg_hue_projection extracted and shared with the trimap
- docs: methodology.md (new), hair_gap_artifacts.md (investigation log)

Verified: cross-check ON reproduces the visually-reviewed B1gate
prototype byte-for-byte on TestImage3; --no-cross-check reproduces the
previous baseline byte-for-byte; pink-bg FixImage1 face untouched.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
2026-07-04 19:52:24 +08:00

178 lines
7.1 KiB
Python

from __future__ import annotations
import numpy as np
from .chroma import bg_hue_projection
from .deps import require_cv2
from .settings import TrimapSettings
def radius_from_ratio(shape: tuple[int, int], ratio: float, minimum: int) -> int:
return max(minimum, int(round(max(shape) * ratio)))
def elliptical_kernel(radius: int) -> np.ndarray:
cv2 = require_cv2()
size = radius * 2 + 1
return cv2.getStructuringElement(cv2.MORPH_ELLIPSE, (size, size))
def generate_trimap(
bg_confidence: np.ndarray, settings: TrimapSettings
) -> tuple[np.ndarray, dict[str, int]]:
cv2 = require_cv2()
shape = bg_confidence.shape
unknown_radius = radius_from_ratio(
shape, settings.unknown_radius_ratio, settings.min_unknown_radius
)
fg_safe_radius = radius_from_ratio(
shape, settings.fg_safe_radius_ratio, settings.min_fg_safe_radius
)
sure_bg = bg_confidence >= settings.sure_bg_threshold
low_bg = bg_confidence <= settings.sure_fg_threshold
bg_u8 = sure_bg.astype(np.uint8)
unknown_band = cv2.dilate(bg_u8, elliptical_kernel(unknown_radius)).astype(bool)
fg_safe = ~cv2.dilate(bg_u8, elliptical_kernel(fg_safe_radius)).astype(bool)
sure_fg = low_bg & fg_safe
trimap = np.full(shape, 128, dtype=np.uint8)
trimap[sure_bg] = 0
trimap[sure_fg] = 255
# Keep a protective unknown band around all sure background, including holes.
trimap[unknown_band & ~sure_bg & ~sure_fg] = 128
stats = {
"sure_bg_pixels": int((trimap == 0).sum()),
"unknown_pixels": int((trimap == 128).sum()),
"sure_fg_pixels": int((trimap == 255).sum()),
"unknown_radius": int(unknown_radius),
"fg_safe_radius": int(fg_safe_radius),
}
return trimap, stats
def fuse_trimap(
seg_mask: np.ndarray, bg_confidence: np.ndarray, settings: TrimapSettings
) -> tuple[np.ndarray, dict[str, int]]:
"""Build a trimap from a semantic subject mask, refined by the chroma key.
Authority split: the segmentation mask decides subject *topology* (it keeps
colour-contaminated hair as foreground and drops see-through holes), while the
chroma key sharpens the flat-background boundary. ViTMatte then refines the
unknown band.
"""
cv2 = require_cv2()
shape = seg_mask.shape
fg_safe_radius = radius_from_ratio(
shape, settings.fg_safe_radius_ratio, settings.min_fg_safe_radius
)
band_radius = max(settings.min_fg_safe_radius, fg_safe_radius // 2)
screen_bg = bg_confidence >= settings.sure_bg_threshold # confident flat background
core = (seg_mask >= settings.seg_core_threshold).astype(np.uint8)
loose = (seg_mask >= settings.seg_loose_threshold).astype(np.uint8)
# A small protective band around the mask boundary lets ViTMatte anti-alias
# crisp edges; hair gets extra width from BiRefNet's own soft region. The band
# is deliberately small so it does not swallow small holes (finger gaps).
kernel = elliptical_kernel(band_radius)
band = cv2.dilate(loose, kernel).astype(bool) & ~cv2.erode(loose, kernel).astype(bool)
# Background follows the mask directly so interior holes stay background, minus
# the boundary band; foreground is the confident subject, never the background.
sure_fg = core.astype(bool) & ~band & ~screen_bg
sure_bg = ((loose == 0) & ~band) | (screen_bg & (loose == 0))
trimap = np.full(shape, 128, dtype=np.uint8)
trimap[sure_bg] = 0
trimap[sure_fg] = 255
stats = {
"sure_bg_pixels": int((trimap == 0).sum()),
"unknown_pixels": int((trimap == 128).sum()),
"sure_fg_pixels": int((trimap == 255).sum()),
"band_radius": int(band_radius),
"fg_safe_radius": int(fg_safe_radius),
"seg_subject_pixels": int((core > 0).sum()),
}
return trimap, stats
def fuse_trimap_directional(
seg_mask: np.ndarray,
bg_confidence: np.ndarray,
lab: np.ndarray,
lab_center: tuple[float, float, float],
settings: TrimapSettings,
) -> tuple[np.ndarray, dict[str, int]]:
"""Three-signal trimap: chroma magnitude + seg + a directional chroma test.
Per pixel (``bgc`` = chroma bg-confidence, ``seg`` = subject confidence)::
BG if bgc >= sure_bg_threshold (clean screen colour)
OR (seg < seg_low AND bgc > sure_fg_threshold) (seg says not-subject,
but never override a pixel chroma is sure is foreground -> keeps
fine wisps the segmenter underestimates)
FG if bgc <= sure_fg_threshold AND seg >= seg_low (clearly non-screen colour)
else (chroma-unknown): if seg >= seg_core_threshold, split by hue --
a pixel *not* displaced toward the background hue (projection <
bg_hue_proj_min: e.g. a neutral white shirt) becomes FG, while one
strongly displaced toward it (e.g. blue between hair strands) is left
unknown for ViTMatte / chroma-suppress. Lower-seg unknowns stay unknown.
A thin unknown band at the segmentation silhouette is preserved so ViTMatte can
anti-alias the boundary (it is deliberately small so interior holes survive).
"""
cv2 = require_cv2()
shape = seg_mask.shape
fg_safe_radius = radius_from_ratio(
shape, settings.fg_safe_radius_ratio, settings.min_fg_safe_radius
)
band_radius = max(settings.min_fg_safe_radius, fg_safe_radius // 2)
chroma_bg = bg_confidence >= settings.sure_bg_threshold
chroma_fg = bg_confidence <= settings.sure_fg_threshold
chroma_unknown = ~chroma_bg & ~chroma_fg
core = seg_mask >= settings.seg_core_threshold
# Directional chroma: positive = colour displaced toward the background hue.
proj = bg_hue_projection(lab, lab_center)
bg_hued = proj >= settings.bg_hue_proj_min
bg = chroma_bg | ((seg_mask < settings.seg_low) & ~chroma_fg)
fg = chroma_fg & (seg_mask >= settings.seg_low)
rule3_fg = chroma_unknown & core & ~bg_hued
if settings.mode == "directional-hard-bg":
# Aggressive variant: hard-remove background-hued pixels instead of leaving
# them unknown. Can eat cool/shadowed white cloth, so it is not the default.
bg = bg | (chroma_unknown & core & bg_hued)
trimap = np.full(shape, 128, dtype=np.uint8)
trimap[bg] = 0
trimap[fg | rule3_fg] = 255
loose = (seg_mask >= settings.seg_loose_threshold).astype(np.uint8)
kernel = elliptical_kernel(band_radius)
band = cv2.dilate(loose, kernel).astype(bool) & ~cv2.erode(loose, kernel).astype(bool)
trimap[band & ~chroma_bg] = 128
stats = {
"sure_bg_pixels": int((trimap == 0).sum()),
"unknown_pixels": int((trimap == 128).sum()),
"sure_fg_pixels": int((trimap == 255).sum()),
"band_radius": int(band_radius),
"rule3_fg_pixels": int(rule3_fg.sum()),
"bg_hued_pixels": int((bg_hued & chroma_unknown & core).sum()),
}
return trimap, stats
def trimap_to_alpha_seed(trimap: np.ndarray, bg_confidence: np.ndarray) -> np.ndarray:
alpha = np.clip(1.0 - bg_confidence, 0.0, 1.0).astype(np.float32)
alpha[trimap == 0] = 0.0
alpha[trimap == 255] = 1.0
return alpha