Spaces:
Running
Running
File size: 11,767 Bytes
0f6f6c1 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 | """Speech bubble detection β find bubble boundaries for accurate text placement.
Uses flood fill on the inpainted image (text already erased) to find
the enclosing speech bubble for each text region. The bubble interior
rectangle is used as the rendering target instead of the tight textline
bounding box, giving translated text much more room to breathe.
"""
import cv2
import numpy as np
from typing import List, Optional
from ..utils import TextBlock, get_logger
logger = get_logger("bubble")
def detect_bubbles(
inpainted_img: np.ndarray,
text_regions: List[TextBlock],
min_bubble_area: int = 800,
max_bubble_ratio: float = 0.3,
padding: int = 12,
) -> List[Optional[np.ndarray]]:
"""
For each text region, detect the enclosing speech bubble.
Returns a list (one per region) of dst_points ``(1, 4, 2)`` int64 arrays
representing the bubble interior rectangle, or ``None`` when no clear
bubble is found (falls back to textline bounding box in the caller).
"""
if inpainted_img.ndim == 3:
gray = cv2.cvtColor(inpainted_img, cv2.COLOR_RGB2GRAY)
else:
gray = inpainted_img.copy()
# Slight blur reduces pixel-level noise that can stop flood fill prematurely
gray = cv2.GaussianBlur(gray, (3, 3), 0)
h, w = gray.shape
img_area = h * w
results: List[Optional[np.ndarray]] = []
for idx, region in enumerate(text_regions):
cx = max(0, min(int(region.center[0]), w - 1))
cy = max(0, min(int(region.center[1]), h - 1))
detected = _detect_single_bubble(
gray, cx, cy, h, w, img_area, region,
min_bubble_area, max_bubble_ratio, padding,
)
rect = None
conf = 0.0
if detected is not None:
rect, conf = detected
region._bubble_confidence = conf
if rect is not None:
bw = int(rect[0, 1, 0] - rect[0, 0, 0])
bh = int(rect[0, 2, 1] - rect[0, 0, 1])
logger.debug(
"Region %d: BUBBLE %dx%d at (%d,%d), conf=%.2f",
idx,
bw,
bh,
int(rect[0, 0, 0]),
int(rect[0, 0, 1]),
conf,
)
else:
logger.debug("Region %d: no bubble detected", idx)
results.append(rect)
# If multiple regions share the same bubble, split the space.
_resolve_overlaps(text_regions, results)
return results
# ββ internals ββββββββββββββββββββββββββββββββββββββββββββββββββββββββ
def _detect_single_bubble(
gray, cx, cy, h, w, img_area, region,
min_area, max_ratio, padding,
) -> Optional[tuple[np.ndarray, float]]:
"""Flood-fill from the text center to find the enclosing bubble."""
tx, ty, tw, th = cv2.boundingRect(region.min_rect[0].astype(np.int32))
text_area = max(1, tw * th)
text_len = len(getattr(region, "translation", "") or region.text or "")
min_area_dyn = max(min_area, int(text_area * 0.9), 300)
max_area_dyn = int(img_area * max_ratio)
best_rect: Optional[np.ndarray] = None
best_score = -1.0
for sx, sy in _candidate_seed_points(cx, cy, tw, th, w, h):
# ββ 1. Flood fill from candidate seed βββββββββββββββββββββββ
bubble_mask, flood_area = _flood_fill(gray, sx, sy)
if flood_area < min_area_dyn:
continue
if flood_area > max_area_dyn:
continue
if flood_area < text_area * 1.05:
continue
contours, _ = cv2.findContours(bubble_mask, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
compactness = 0.0
if contours:
cnt = max(contours, key=cv2.contourArea)
hull_area = cv2.contourArea(cv2.convexHull(cnt))
if hull_area > 0:
compactness = cv2.contourArea(cnt) / hull_area
# Relax compactness threshold for dense text to avoid over-rejection.
density = text_len / max(np.sqrt(max(flood_area, 1)), 1.0)
compact_thresh = max(0.28, min(0.42, 0.40 - 0.10 * density))
if compactness > 0 and compactness < compact_thresh:
continue
# ββ 2. Erode to safe interior while keeping selected seed inside ββββ
adaptive_padding = _adaptive_padding(flood_area, region, tw, th, padding)
eroded = _erode_mask(bubble_mask, sx, sy, adaptive_padding)
if eroded is None:
continue
num_labels, labels = cv2.connectedComponents(eroded)
label_at_seed = labels[sy, sx]
if label_at_seed == 0:
continue
component = (labels == label_at_seed).astype(np.uint8) * 255
bx, by, bw, bh = cv2.boundingRect(component)
min_dim = max(18, int(min(tw, th) * 0.55))
if bw < min_dim or bh < min_dim:
continue
if region.horizontal and bw < tw * 0.7:
continue
score = _bubble_confidence_score(flood_area, text_area, compactness, bw, bh, tw, th)
if score > best_score:
best_score = score
best_rect = _rect_to_dst(bx, by, bw, bh)
if best_rect is None:
return None
return best_rect, float(max(0.0, min(1.0, best_score)))
def _candidate_seed_points(cx: int, cy: int, tw: int, th: int, w: int, h: int) -> list[tuple[int, int]]:
delta = max(2, int(round(min(tw, th) * 0.10)))
pts = [
(cx, cy),
(cx + delta, cy),
(cx - delta, cy),
(cx, cy + delta),
(cx, cy - delta),
]
uniq = set()
out = []
for x, y in pts:
sx = max(0, min(int(x), w - 1))
sy = max(0, min(int(y), h - 1))
key = (sx, sy)
if key not in uniq:
uniq.add(key)
out.append(key)
return out
def _flood_fill(gray: np.ndarray, sx: int, sy: int) -> tuple[np.ndarray, int]:
flood_mask = np.zeros((gray.shape[0] + 2, gray.shape[1] + 2), np.uint8)
gray_copy = gray.copy()
cv2.floodFill(
gray_copy,
flood_mask,
(int(sx), int(sy)),
newVal=0,
loDiff=(35,),
upDiff=(35,),
flags=cv2.FLOODFILL_MASK_ONLY | (255 << 8),
)
bubble_mask = flood_mask[1:-1, 1:-1]
flood_area = int(np.sum(bubble_mask > 0))
return bubble_mask, flood_area
def _bubble_confidence_score(
flood_area: int,
text_area: int,
compactness: float,
bw: int,
bh: int,
tw: int,
th: int,
) -> float:
area_ratio = flood_area / max(text_area, 1)
area_term = max(0.0, min(1.0, (area_ratio - 1.0) / 4.0))
compact_term = max(0.0, min(1.0, compactness))
width_fit = max(0.0, min(1.0, bw / max(tw, 1)))
height_fit = max(0.0, min(1.0, bh / max(th, 1)))
fit_term = 0.5 * width_fit + 0.5 * height_fit
return 0.45 * area_term + 0.30 * compact_term + 0.25 * fit_term
def _erode_mask(mask, cx, cy, padding):
"""Erode *mask*, reducing padding until *center* is still inside."""
h, w = mask.shape
if cy < 0 or cy >= h or cx < 0 or cx >= w:
return None
for p in range(padding, 1, -2):
kernel = np.ones((p * 2 + 1, p * 2 + 1), np.uint8)
eroded = cv2.erode(mask, kernel)
if eroded[cy, cx] > 0:
return eroded
# Minimal / no erosion
if mask[cy, cx] > 0:
return mask
return None
def _adaptive_padding(
flood_area: int,
region: TextBlock,
text_w: int,
text_h: int,
base_padding: int,
) -> int:
"""Compute erosion padding from bubble size and text density.
Uses a smooth formula rather than fixed area classes:
- Larger bubbles get more padding.
- Longer/denser text gets less padding to avoid cramped rendering.
- Preserves caller-provided ``base_padding`` as a soft prior.
"""
bubble_dim = max(1.0, float(np.sqrt(max(flood_area, 1))))
text_len = len(getattr(region, "translation", "") or region.text or "")
# Size-driven padding component (smooth growth with bubble dimension)
size_padding = 0.06 * bubble_dim
# Text density proxy: higher density => less interior erosion
text_density = text_len / max(bubble_dim, 1.0)
density_factor = max(0.72, min(1.12, 1.06 - 0.16 * text_density))
# Blend caller default with adaptive value for backward compatibility
blended = (0.45 * float(base_padding) + 0.55 * size_padding) * density_factor
# Prevent erosion from consuming tiny bubbles
upper_bound = max(3, int(min(text_w, text_h) * 0.22))
return int(max(3, min(24, min(upper_bound, round(blended)))))
def _resolve_overlaps(text_regions, bubble_rects):
"""Split shared bubbles among multiple text regions."""
n = len(bubble_rects)
for i in range(n):
if bubble_rects[i] is None:
continue
for j in range(i + 1, n):
if bubble_rects[j] is None:
continue
r1 = cv2.boundingRect(bubble_rects[i][0].astype(np.int32))
r2 = cv2.boundingRect(bubble_rects[j][0].astype(np.int32))
if _rect_iou(r1, r2) < 0.5:
continue
_split_shared_bubble(text_regions, bubble_rects, i, j)
def _rect_iou(r1, r2):
x1 = max(r1[0], r2[0])
y1 = max(r1[1], r2[1])
x2 = min(r1[0] + r1[2], r2[0] + r2[2])
y2 = min(r1[1] + r1[3], r2[1] + r2[3])
inter = max(0, x2 - x1) * max(0, y2 - y1)
union = r1[2] * r1[3] + r2[2] * r2[3] - inter
return inter / union if union > 0 else 0.0
def _split_shared_bubble(text_regions, bubble_rects, i, j):
"""Split shared bubble along dominant region-center axis."""
ri = bubble_rects[i][0]
bx = int(ri[0, 0])
by = int(ri[0, 1])
bw = int(ri[1, 0] - bx)
bh = int(ri[2, 1] - by)
cy_i = float(text_regions[i].center[1])
cy_j = float(text_regions[j].center[1])
cx_i = float(text_regions[i].center[0])
cx_j = float(text_regions[j].center[0])
if abs(cx_i - cx_j) > abs(cy_i - cy_j):
_split_horizontally(text_regions, bubble_rects, i, j, bx, by, bw, bh, cx_i, cx_j)
else:
_split_vertically(text_regions, bubble_rects, i, j, bx, by, bw, bh, cy_i, cy_j)
def _split_vertically(text_regions, bubble_rects, i, j, bx, by, bw, bh, cy_i, cy_j):
"""Split a shared bubble between two regions by Y axis."""
split_y = int((cy_i + cy_j) / 2)
split_y = max(by + 10, min(split_y, by + bh - 10))
if cy_i <= cy_j:
h_top = split_y - by
h_bot = by + bh - split_y
bubble_rects[i] = _rect_to_dst(bx, by, bw, h_top)
bubble_rects[j] = _rect_to_dst(bx, split_y, bw, h_bot)
else:
h_top = split_y - by
h_bot = by + bh - split_y
bubble_rects[j] = _rect_to_dst(bx, by, bw, h_top)
bubble_rects[i] = _rect_to_dst(bx, split_y, bw, h_bot)
def _split_horizontally(text_regions, bubble_rects, i, j, bx, by, bw, bh, cx_i, cx_j):
"""Split a shared bubble between two regions by X axis."""
split_x = int((cx_i + cx_j) / 2)
split_x = max(bx + 10, min(split_x, bx + bw - 10))
if cx_i <= cx_j:
w_left = split_x - bx
w_right = bx + bw - split_x
bubble_rects[i] = _rect_to_dst(bx, by, w_left, bh)
bubble_rects[j] = _rect_to_dst(split_x, by, w_right, bh)
else:
w_left = split_x - bx
w_right = bx + bw - split_x
bubble_rects[j] = _rect_to_dst(bx, by, w_left, bh)
bubble_rects[i] = _rect_to_dst(split_x, by, w_right, bh)
def _rect_to_dst(x, y, w, h):
"""Pack (x, y, w, h) into a ``(1, 4, 2)`` int64 dst_points array."""
return np.array(
[[[x, y], [x + w, y], [x + w, y + h], [x, y + h]]],
dtype=np.int64,
)
|