File size: 11,388 Bytes
0fcd8ae |
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 |
"""
Utility functions for surgical instrument classification
"""
import cv2
import numpy as np
from skimage.feature.texture import graycomatrix, graycoprops
from skimage.feature import local_binary_pattern, hog
from sklearn.decomposition import PCA
from sklearn.svm import SVC
from sklearn.model_selection import train_test_split
from sklearn.metrics import accuracy_score, f1_score
import pywt
def preprocess_image(image):
"""
Apply CLAHE preprocessing for better contrast
(Contrast Limited Adaptive Historam Equalization)
"""
# Convert to LAB color space (basically separating lightness, L, from color info)
lab = cv2.cvtColor(image, cv2.COLOR_BGR2LAB)
l, a, b = cv2.split(lab) #this enhances constrast between colors
# Apply CLAHE to L channel
clahe = cv2.createCLAHE(clipLimit=2.0, tileGridSize=(8,8)) #split into a 8x8 grid and performs the contrast enhancement to the smaller regions instead of full image
l = clahe.apply(l)
# Merge and convert back
enhanced = cv2.merge([l, a, b]) #merge the contrast channel with the other two (A,B)
enhanced = cv2.cvtColor(enhanced, cv2.COLOR_LAB2BGR) #go back to BGR so it can be used later on
return enhanced
#this is the same as baseline code, well working so let's keep it
#it basically computes normalized color histograms for the classic three channels
def rgb_histogram(image, bins=256):
"""Extract RGB histogram features"""
hist_features = []
for i in range(3): # RGB Channels
hist, _ = np.histogram(image[:, :, i], bins=bins, range=(0, 256), density=True)
hist_features.append(hist)
return np.concatenate(hist_features)
def hu_moments(image):
"""Extract Hu moment features, takes BGR format in input
basically provides shape description that are consistent
wrt to position, size and rotation"""
gray = cv2.cvtColor(image, cv2.COLOR_BGR2GRAY) #turn to greyscale (works in 1 channel)
moments = cv2.moments(gray)
hu_moments = cv2.HuMoments(moments).flatten()
return hu_moments
def glcm_features(image, distances=[1], angles=[0], levels=256, symmetric=True, normed=True):
"""Extract GLCM texture features,
captures texture info considering spatial
relationship between pixel intensities. works well with RGB and hu"""
gray = cv2.cvtColor(image, cv2.COLOR_BGR2GRAY)
glcm = graycomatrix(gray, distances=distances, angles=angles, levels=levels,
symmetric=symmetric, normed=normed)
contrast = graycoprops(glcm, 'contrast').flatten()
dissimilarity = graycoprops(glcm, 'dissimilarity').flatten()
homogeneity = graycoprops(glcm, 'homogeneity').flatten()
energy = graycoprops(glcm, 'energy').flatten()
correlation = graycoprops(glcm, 'correlation').flatten()
asm = graycoprops(glcm, 'ASM').flatten()
return np.concatenate([contrast, dissimilarity, homogeneity, energy, correlation, asm])
def local_binary_pattern_features(image, P=8, R=1):
"""Extract Local Binary Pattern features, useful for light changes
combined with rgb, hu and glcm"""
gray = cv2.cvtColor(image, cv2.COLOR_BGR2GRAY)
lbp = local_binary_pattern(gray, P, R, method='uniform')
(hist, _) = np.histogram(lbp.ravel(), bins=np.arange(0, P + 3),
range=(0, P + 2), density=True)
return hist #feature vector representing the texture of the image
def hog_features(image, orientations=12, pixels_per_cell=(16, 16), cells_per_block=(2, 2)):
"""
Extract HOG (Histogram of Oriented Gradients) features
Great for capturing shape and edge information in surgical instruments
"""
gray = cv2.cvtColor(image, cv2.COLOR_BGR2GRAY)
# Resize to standard size for consistency
gray_resized = cv2.resize(gray, (256, 256)) #we could try using 256 here and 16,16 cells per block
hog_features_vector = hog(
gray_resized,
orientations=orientations,
pixels_per_cell=pixels_per_cell,
cells_per_block=cells_per_block,
block_norm='L2-Hys',
feature_vector=True
)
return hog_features_vector #Returns a vector capturing local edge
#directions and shape information, useful for detecting instruments,
#objects, or structural patterns.
def luv_histogram(image, bins=32): #instead of bgr it uses lightness and chromatic components
"""
Extract histogram in LUV color space
LUV is perceptually uniform and better for underwater/surgical imaging
"""
luv = cv2.cvtColor(image, cv2.COLOR_BGR2LUV)
hist_features = []
for i in range(3):
hist, _ = np.histogram(luv[:, :, i], bins=bins, range=(0, 256), density=True)
hist_features.append(hist)
return np.concatenate(hist_features)
def gabor_features(image, frequencies=[0.1, 0.2, 0.3],
orientations=[0, 45, 90, 135]):
"""
Extract Gabor filter features (gabor kernels)
texture orientation that deals well with different scales and diff orientation
"""
gray = cv2.cvtColor(image, cv2.COLOR_BGR2GRAY) # uses intensity and not color
features = []
for freq in frequencies:
for theta in orientations:
theta_rad = theta * np.pi / 180
kernel = cv2.getGaborKernel((21, 21), 5, theta_rad,
10.0/freq, 0.5, 0)
filtered = cv2.filter2D(gray, cv2.CV_32F, kernel)
features.append(np.mean(filtered))
features.append(np.std(filtered))
return np.array(features)
def wavelet_features(image, wavelet='db4', levels=3):
# try out multi-orientation wavelets (e.g., sym8, coif)
"""
multi-scale wavelet feature extractor
focus on texture + edges.
This function converts an image to grayscale,
performs a multi-level 2D wavelet decomposition,
and extracts texture features from the approximation and detail sub-bands.
It summarizes each band using statistics like mean, standard deviation, and peak magnitude.
The result is a numerical feature vector capturing multi-scale texture and edge
information for tasks like surgical tool detection.
"""
# Convert to grayscale float32 in [0,1]
gray = cv2.cvtColor(image, cv2.COLOR_BGR2GRAY).astype(np.float32) / 255.0
# Ensure the decomposition level is valid for this image
max_level = pywt.dwt_max_level(min(gray.shape), pywt.Wavelet(wavelet).dec_len)
levels = min(levels, max_level)
coeffs = pywt.wavedec2(gray, wavelet=wavelet, level=levels)
features = []
#Approximation Coefficients (LL)
LL = coeffs[0]
LL_abs = np.abs(LL)
features.extend([
LL.mean(),
LL.std(),
LL_abs.max(),
LL_abs.mean(),
])
# Detail Coefficients for each level: (LH, HL, HH)
for (LH, HL, HH) in coeffs[1:]:
for band in (LH, HL, HH):
band_abs = np.abs(band)
features.extend([
band_abs.mean(), # energy-like texture measure
band_abs.std(), # variation in texture
band_abs.max(), # strongest directional edge
np.percentile(band_abs, 95), # robust peak measure
])
return np.array(features, dtype=np.float32)
def extract_features_from_image(image):
"""
Extract enhanced features from image
Uses baseline features + HOG + LUV histogram + Gabor for better performance
Args:
image: Input image (BGR format from cv2.imread)
Returns:
Feature vector as numpy array
"""
# Preprocess image first
image = preprocess_image(image)
# Baseline features
hist_features = rgb_histogram(image)
hu_features = hu_moments(image)
glcm_features_vector = glcm_features(image)
lbp_features = local_binary_pattern_features(image)
# Enhanced features that add discriminative power for complex images
hog_feat = hog_features(image)
luv_hist = luv_histogram(image)
gabor_feat = gabor_features(image)
wavelet_feat = wavelet_features(image)
# Concatenate all features (produces a single vector)
image_features = np.concatenate([
hist_features,
hu_features,
glcm_features_vector,
lbp_features,
hog_feat,
luv_hist,
gabor_feat,
wavelet_feat
])
return image_features # comprehensive numerical representation of the imag
def fit_pca_transformer(data, num_components):
"""
Fit a PCA transformer on training data
Args:
data: Training data (n_samples, n_features)
num_components: Number of PCA components to keep
Returns:
pca_params: Dictionary containing PCA parameters
data_reduced: PCA-transformed data
"""
# Standardize the data
mean = np.mean(data, axis=0)
std = np.std(data, axis=0)
# Avoid division by zero
std[std == 0] = 1.0
data_standardized = (data - mean) / std
# Fit PCA using sklearn
pca_model = PCA(n_components=num_components)
data_reduced = pca_model.fit_transform(data_standardized)
# Create params dictionary
pca_params = {
'pca_model': pca_model,
'mean': mean,
'std': std,
'num_components': num_components,
'feature_dim': data.shape[1],
'explained_variance_ratio': pca_model.explained_variance_ratio_,
'cumulative_variance': np.cumsum(pca_model.explained_variance_ratio_)
}
return pca_params, data_reduced
def apply_pca_transform(data, pca_params):
"""
Apply saved PCA transformation to new data
CRITICAL: This uses the saved mean/std/PCA from training
Args:
data: New data to transform (n_samples, n_features)
pca_params: Dictionary from fit_pca_transformer
Returns:
Transformed data
"""
# Standardize using training mean/std
data_standardized = (data - pca_params['mean']) / pca_params['std']
# Apply PCA transformation
# Projects new data onto the same principal components computed from training data
data_reduced = pca_params['pca_model'].transform(data_standardized)
return data_reduced
def train_svm_model(features, labels, kernel='rbf', C=1.0):
"""
Train an SVM model on ALL available data (no train/test split)
Args:
features: Feature matrix (n_samples, n_features)
labels: Label array (n_samples,)
kernel: SVM kernel type ('linear', 'rbf', 'poly', 'sigmoid')
C: Regularization parameter (smaller = more regularization)
Returns:
Dictionary containing model and metrics
"""
# Check if labels are one-hot encoded
if labels.ndim > 1 and labels.shape[1] > 1:
labels = np.argmax(labels, axis=1)
# Train SVM on ALL data
svm_model = SVC(kernel=kernel, C=C, random_state=56)
svm_model.fit(features, labels)
results = {
'model': svm_model
}
return results |