File size: 6,986 Bytes
5c22e5c | 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 | """
features.py — 120-dimensional handcrafted CV feature extraction.
Pipeline (matches training notebook exactly):
preprocess(img) → get_mask(img) → extract_features(img, mask)
Feature vector layout:
[0:24] GLCM texture (6 props × 4 angles)
[24:50] LBP texture (26-bin uniform histogram, P=24 R=3)
[50:74] Gabor texture (4 freqs × 3 orientations, mean+std)
[74:106] Colour histogram (16H + 8S + 8V, normalised, within mask)
[106:115] Colour moments (mean, std, skewness per HSV channel)
[115:120] ABCD morphology (asymmetry, compactness, cvar, diam, elong)
Total: 120 dimensions
"""
import numpy as np
import cv2
from skimage.feature import graycomatrix, graycoprops, local_binary_pattern
# ── Preprocessing ──────────────────────────────────────────────────────────────
def preprocess(img_bgr: np.ndarray, size: tuple = (256, 256)) -> np.ndarray:
"""Hair removal (black-hat) → Gaussian blur → CLAHE → resize."""
gray = cv2.cvtColor(img_bgr, cv2.COLOR_BGR2GRAY)
kernel = cv2.getStructuringElement(cv2.MORPH_RECT, (17, 17))
bhat = cv2.morphologyEx(gray, cv2.MORPH_BLACKHAT, kernel)
_, hmask = cv2.threshold(bhat, 10, 255, cv2.THRESH_BINARY)
img_bgr = cv2.inpaint(img_bgr, hmask, 3, cv2.INPAINT_TELEA)
img_bgr = cv2.GaussianBlur(img_bgr, (5, 5), 0)
lab = cv2.cvtColor(img_bgr, cv2.COLOR_BGR2LAB)
clahe = cv2.createCLAHE(clipLimit=2.0, tileGridSize=(8, 8))
lab[:, :, 0] = clahe.apply(lab[:, :, 0])
img_bgr = cv2.cvtColor(lab, cv2.COLOR_LAB2BGR)
return cv2.resize(img_bgr, size, interpolation=cv2.INTER_AREA)
def get_mask(img_bgr: np.ndarray) -> np.ndarray:
"""Otsu thresholding + morphological refinement to isolate lesion."""
gray = cv2.cvtColor(img_bgr, cv2.COLOR_BGR2GRAY)
blur = cv2.GaussianBlur(gray, (7, 7), 0)
_, mask = cv2.threshold(blur, 0, 255, cv2.THRESH_BINARY_INV + cv2.THRESH_OTSU)
k = cv2.getStructuringElement(cv2.MORPH_ELLIPSE, (15, 15))
mask = cv2.morphologyEx(mask, cv2.MORPH_CLOSE, k)
mask = cv2.morphologyEx(
mask, cv2.MORPH_OPEN,
cv2.getStructuringElement(cv2.MORPH_ELLIPSE, (5, 5))
)
return mask
# ── Feature groups ─────────────────────────────────────────────────────────────
def feat_glcm(gray: np.ndarray) -> np.ndarray:
"""24-dim GLCM at 4 angles × 6 properties."""
glcm = graycomatrix(
gray, distances=[1],
angles=[0, np.pi / 4, np.pi / 2, 3 * np.pi / 4],
levels=256, symmetric=True, normed=True
)
props = ["contrast", "correlation", "energy", "homogeneity", "dissimilarity", "ASM"]
return np.concatenate([graycoprops(glcm, p).flatten() for p in props])
def feat_lbp(gray: np.ndarray) -> np.ndarray:
"""26-dim LBP histogram — rotation-invariant uniform, P=24 R=3."""
lbp = local_binary_pattern(gray, P=24, R=3, method="uniform")
hist, _ = np.histogram(lbp.ravel(), bins=26, range=(0, 26), density=True)
return hist.astype(np.float32)
def feat_gabor(gray: np.ndarray) -> np.ndarray:
"""24-dim Gabor responses at 4 scales × 3 orientations (mean + std each)."""
feats = []
for freq in [0.1, 0.2, 0.3, 0.4]:
for theta in [0, np.pi / 3, 2 * np.pi / 3]:
kernel = cv2.getGaborKernel(
(21, 21), sigma=4.0, theta=theta,
lambd=1.0 / freq, gamma=0.5, psi=0
)
resp = cv2.filter2D(gray.astype(np.float32), cv2.CV_32F, kernel)
feats.extend([resp.mean(), resp.std()])
return np.array(feats, dtype=np.float32)
def feat_colour(img_bgr: np.ndarray, mask: np.ndarray) -> np.ndarray:
"""32-dim normalised HSV histogram within lesion mask (16H + 8S + 8V)."""
hsv = cv2.cvtColor(img_bgr, cv2.COLOR_BGR2HSV)
m = (mask > 0).astype(np.uint8) * 255
h = cv2.calcHist([hsv], [0], m, [16], [0, 180]).flatten()
s = cv2.calcHist([hsv], [1], m, [8], [0, 256]).flatten()
v = cv2.calcHist([hsv], [2], m, [8], [0, 256]).flatten()
norm = lambda x: x / (x.sum() + 1e-8)
return np.concatenate([norm(h), norm(s), norm(v)]).astype(np.float32)
def feat_colour_moments(img_bgr: np.ndarray, mask: np.ndarray) -> np.ndarray:
"""9-dim colour moments (mean, std, skewness) per HSV channel."""
hsv = cv2.cvtColor(img_bgr, cv2.COLOR_BGR2HSV)
feats = []
for ch in range(3):
pixels = hsv[:, :, ch][mask > 0].astype(float)
if pixels.size == 0:
feats.extend([0.0, 0.0, 0.0])
continue
mu = pixels.mean()
sigma = pixels.std() + 1e-8
skew = float(np.mean(((pixels - mu) / sigma) ** 3))
feats.extend([mu, sigma, skew])
return np.array(feats, dtype=np.float32)
def feat_abcd(mask: np.ndarray, hsv: np.ndarray) -> np.ndarray:
"""5-dim ABCD dermoscopy features (asymmetry, compactness, cvar, diam, elong)."""
cnts, _ = cv2.findContours(mask, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
if not cnts:
return np.zeros(5, dtype=np.float32)
cnt = max(cnts, key=cv2.contourArea)
area = cv2.contourArea(cnt) + 1e-6
perim = cv2.arcLength(cnt, True) + 1e-6
if len(cnt) >= 5:
axes_sorted = sorted(cv2.fitEllipse(cnt)[1])
asym = axes_sorted[0] / (axes_sorted[1] + 1e-6)
else:
asym = 1.0
comp = (4 * np.pi * area) / (perim ** 2)
hue_pixels = hsv[:, :, 0][mask > 0]
cvar = float(hue_pixels.std()) if hue_pixels.size else 0.0
diam = np.sqrt(4 * area / np.pi)
elong = perim / (2 * np.sqrt(np.pi * area) + 1e-6)
return np.array([asym, comp, cvar, diam, elong], dtype=np.float32)
# ── Full 120D extraction ───────────────────────────────────────────────────────
def extract_features(img_bgr: np.ndarray) -> np.ndarray:
"""
Full pipeline: preprocess → mask → 120D feature vector.
Input : BGR image (any size)
Output: float32 ndarray of shape (120,)
"""
proc = preprocess(img_bgr) # 256×256 BGR
mask = get_mask(proc) # binary mask
gray = cv2.cvtColor(proc, cv2.COLOR_BGR2GRAY)
hsv = cv2.cvtColor(proc, cv2.COLOR_BGR2HSV)
vec = np.concatenate([
feat_glcm(gray), # 24
feat_lbp(gray), # 26
feat_gabor(gray), # 24
feat_colour(proc, mask), # 32
feat_colour_moments(proc, mask), # 9
feat_abcd(mask, hsv), # 5
]).astype(np.float32)
assert vec.shape == (120,), f"Expected 120 dims, got {vec.shape}"
return vec
|