File size: 8,050 Bytes
b5b458d |
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 |
"""
Feature extractors for PosterSentry.
Two feature channels:
1. Visual features — image-level statistics (color, edges, FFT, whitespace)
2. PDF structural features — page geometry, text blocks, font diversity
Both are cheap to compute (no GPU needed), providing strong priors that
complement the text embedding from model2vec.
"""
import logging
from pathlib import Path
from typing import Dict, List, Optional, Tuple
import numpy as np
logger = logging.getLogger(__name__)
# ── Visual Feature Extractor ────────────────────────────────────
VISUAL_FEATURE_NAMES = [
"img_width",
"img_height",
"img_aspect_ratio",
"mean_r", "mean_g", "mean_b",
"std_r", "std_g", "std_b",
"local_contrast",
"color_diversity",
"edge_density",
"spatial_complexity",
"white_space_ratio",
"high_contrast_ratio",
]
N_VISUAL_FEATURES = len(VISUAL_FEATURE_NAMES)
class VisualFeatureExtractor:
"""Extract visual features from rendered PDF pages."""
FEATURE_NAMES = VISUAL_FEATURE_NAMES
def __init__(self, target_size: Tuple[int, int] = (256, 256)):
self.target_size = target_size
def pdf_to_image(self, pdf_path: str, dpi: int = 72) -> Optional[np.ndarray]:
"""Render first page of PDF to RGB numpy array."""
try:
import fitz
doc = fitz.open(pdf_path)
if len(doc) == 0:
doc.close()
return None
page = doc[0]
mat = fitz.Matrix(dpi / 72, dpi / 72)
pix = page.get_pixmap(matrix=mat)
img = np.frombuffer(pix.samples, dtype=np.uint8).reshape(pix.height, pix.width, pix.n)
if pix.n == 4:
img = img[:, :, :3]
elif pix.n == 1:
img = np.stack([img[:, :, 0]] * 3, axis=-1)
doc.close()
return img
except Exception as e:
logger.debug(f"PDF to image failed: {e}")
return None
def extract(self, image: np.ndarray) -> Dict[str, float]:
"""Extract 15 visual features from an RGB image."""
feats = {n: 0.0 for n in self.FEATURE_NAMES}
try:
from PIL import Image as PILImage
h, w = image.shape[:2]
feats["img_width"] = float(w)
feats["img_height"] = float(h)
feats["img_aspect_ratio"] = w / h if h > 0 else 0.0
pil = PILImage.fromarray(image).resize(self.target_size, PILImage.Resampling.BILINEAR)
resized = np.array(pil)
for i, ch in enumerate(["r", "g", "b"]):
feats[f"mean_{ch}"] = float(np.mean(resized[:, :, i]))
feats[f"std_{ch}"] = float(np.std(resized[:, :, i]))
gray = np.mean(resized, axis=2)
feats["local_contrast"] = float(np.std(gray))
# Color diversity (unique quantized colors in 32x32 thumbnail)
small = np.array(pil.resize((32, 32)))
quantized = (small // 32).astype(np.uint8)
unique_colors = len(np.unique(quantized.reshape(-1, 3), axis=0))
feats["color_diversity"] = unique_colors / 512.0
# Edge density
gy = np.abs(np.diff(gray, axis=0))
gx = np.abs(np.diff(gray, axis=1))
feats["edge_density"] = float(np.mean(gy) + np.mean(gx)) / 255.0
# Spatial complexity (high-freq ratio via FFT)
fft = np.fft.fft2(gray)
fft_shift = np.fft.fftshift(fft)
mag = np.abs(fft_shift)
ch, cw = mag.shape[0] // 2, mag.shape[1] // 2
radius = min(mag.shape) // 4
y, x = np.ogrid[:mag.shape[0], :mag.shape[1]]
center_mask = ((y - ch) ** 2 + (x - cw) ** 2) <= radius ** 2
total_e = np.sum(mag ** 2)
low_e = np.sum(mag[center_mask] ** 2)
feats["spatial_complexity"] = 1.0 - (low_e / total_e) if total_e > 0 else 0.0
# White space ratio
white_px = np.sum(np.all(resized > 240, axis=2))
feats["white_space_ratio"] = white_px / (self.target_size[0] * self.target_size[1])
# High contrast ratio (very dark + very bright pixels)
feats["high_contrast_ratio"] = float(np.sum(gray < 50) + np.sum(gray > 240)) / gray.size
except Exception as e:
logger.debug(f"Visual feature extraction failed: {e}")
return feats
def to_vector(self, feats: Dict[str, float]) -> np.ndarray:
return np.array([feats.get(n, 0.0) for n in self.FEATURE_NAMES], dtype="float32")
# ── PDF Structural Feature Extractor ────────────────────────────
STRUCTURAL_FEATURE_NAMES = [
"page_count",
"page_width_pt",
"page_height_pt",
"page_aspect_ratio",
"page_area_sqin",
"is_landscape",
"text_block_count",
"font_count",
"avg_font_size",
"font_size_variance",
"title_score",
"text_density",
"line_count",
"file_size_kb",
"size_per_page_kb",
]
N_STRUCTURAL_FEATURES = len(STRUCTURAL_FEATURE_NAMES)
class PDFStructuralExtractor:
"""Extract structural features from PDF layout."""
FEATURE_NAMES = STRUCTURAL_FEATURE_NAMES
def extract(self, pdf_path: str) -> Dict[str, float]:
"""Extract 15 structural features from a PDF."""
feats = {n: 0.0 for n in self.FEATURE_NAMES}
try:
import fitz
path = Path(pdf_path)
doc = fitz.open(str(path))
if len(doc) == 0:
doc.close()
return feats
feats["page_count"] = float(len(doc))
feats["file_size_kb"] = path.stat().st_size / 1024.0
feats["size_per_page_kb"] = feats["file_size_kb"] / max(len(doc), 1)
page = doc[0]
rect = page.rect
feats["page_width_pt"] = rect.width
feats["page_height_pt"] = rect.height
feats["page_aspect_ratio"] = rect.width / rect.height if rect.height > 0 else 0.0
feats["page_area_sqin"] = (rect.width / 72.0) * (rect.height / 72.0)
feats["is_landscape"] = float(rect.width > rect.height)
# Text blocks
blocks = page.get_text("dict")["blocks"]
text_blocks = [b for b in blocks if b.get("type") == 0]
feats["text_block_count"] = float(len(text_blocks))
if text_blocks:
heights = [b["bbox"][3] - b["bbox"][1] for b in text_blocks]
widths = [b["bbox"][2] - b["bbox"][0] for b in text_blocks]
total_area = sum(h * w for h, w in zip(heights, widths))
page_area = rect.width * rect.height
feats["text_density"] = total_area / page_area if page_area > 0 else 0.0
# Font statistics
fonts = set()
font_sizes = []
line_count = 0
for block in text_blocks:
for line in block.get("lines", []):
line_count += 1
for span in line.get("spans", []):
fonts.add(span.get("font", ""))
sz = span.get("size", 0)
if sz > 0:
font_sizes.append(sz)
feats["font_count"] = float(len(fonts))
feats["line_count"] = float(line_count)
if font_sizes:
feats["avg_font_size"] = float(np.mean(font_sizes))
feats["font_size_variance"] = float(np.var(font_sizes)) if len(font_sizes) > 1 else 0.0
feats["title_score"] = max(font_sizes) / (np.mean(font_sizes) + 1.0)
doc.close()
except Exception as e:
logger.debug(f"PDF structural extraction failed: {e}")
return feats
def to_vector(self, feats: Dict[str, float]) -> np.ndarray:
return np.array([feats.get(n, 0.0) for n in self.FEATURE_NAMES], dtype="float32")
|