import image_operations as operations
import feature_extraction as extraction
from skimage import feature, color, exposure
import feature_validation as validation
training_images, training_labels, training_classes = data_loading.loadTrainingImagesPoleNumbersAndClasses(
)
size = 100
print("resizing...")
resized = util.loading_map(lambda x: operations.cropAndResize(x, 0, size),
training_images)
print("hsv...")
hsv = util.loading_map(color.rgb2hsv, resized)
print("grayscaling...")
grayscaled = util.loading_map(color.rgb2gray, resized)
print("colors")
colors = util.loading_map(
lambda x: extraction.split_image_features(
extraction.calculateColorFeatures, 7, x), hsv)
n_folds = 5
print("evaluating colors")
model = DistanceModel()
#from sklearn.linear_model import LogisticRegression
#model = LogisticRegression(solver = 'lbfgs', multi_class = 'multinomial')
validation.validate_feature(colors, training_labels, training_classes, model,
n_folds, False, False, True, True)
print('\a')
model = Pipeline([("standard scaler", StandardScaler()),
("logistic regression",
LogisticRegression(solver='lbfgs',
multi_class='multinomial'))])
for cpb in range(2, 11):
ppc = int(100 / cpb)
print("hog_8_", ppc, "_", cpb, " features")
hog = util.loading_map(
lambda x: feature.hog(x,
orientations=8,
pixels_per_cell=(ppc, ppc),
cells_per_block=(cpb, cpb),
normalise=True), grayscaled)
print(numpy.shape(hog))
validation.validate_feature(hog, labels, classes, model, n_folds, False,
False, True, True)
#hog_8_12_8 is optimal, optimizing orientations:
for orientations in range(10, 11):
print("hog_", orientations, "_12_8 features")
hog = util.loading_map(
lambda x: feature.hog(x,
orientations=orientations,
pixels_per_cell=(12, 12),
cells_per_block=(8, 8),
normalise=True), grayscaled)
print(numpy.shape(hog))
validation.validate_feature(hog, labels, classes, model, n_folds, False,
False, True, True)
#retry optimizing cpb using 6 orientations
model = Pipeline([
("standard scaler", StandardScaler()),
#("Robust Scaler", RobustScaler()),
#("principal component analysis", PCA(500)), #<- appears to reduce efficiency
#("lda projection", lda.LDA(n_components = 80)),
#("gaussian random projection", random_projection.GaussianRandomProjection(n_components = 100)),
#("sparse random projection", random_projection.SparseRandomProjection(n_components = 500)),
("logistic regression",
LogisticRegression(solver='lbfgs', multi_class='multinomial'))
#("Random forest classifier", RandomForestClassifier(n_estimators = 200)),
#("svm", svm.SVC(kernel = "sigmoid", C = 1000, gamma = 0.0001, probability = True))
])
print("Evaluating brightness features")
validation.validate_feature(brightness, labels, classes, model, n_folds, False,
False, True)
print('\a')
print("Evaluating hsv features")
validation.validate_feature(hsv_features, labels, classes, model, n_folds,
False, False, True, True)
print("Evaluating luv features")
validation.validate_feature(luv_features, labels, classes, model, n_folds,
False, False, True, True)
print("Evaluating hed features")
validation.validate_feature(hed_features, labels, classes, model, n_folds,
False, False, True, True)
print("Evaluating cie features")
validation.validate_feature(cie_features, labels, classes, model, n_folds,
False, False, True, True)
print("Evaluating HOG features")
#validation.validate_feature(hog, labels, classes, model, n_folds, False, False, True, True)
hsv = util.loading_map(color.rgb2hsv, resized)
print("luv...")
luv = util.loading_map(color.rgb2luv, resized)
model = Pipeline([("standard scaler", StandardScaler()),
("logistic regression",
LogisticRegression(solver='lbfgs',
multi_class='multinomial'))])
n_folds = 10
for i in range(1, 14, 2):
print("rgb", i)
rgb_f = util.loading_map(lambda x: extraction.pixel_features(x, i),
resized)
validation.validate_feature(rgb_f, labels, classes, model, n_folds, False,
False, True, True)
for i in range(1, 14, 2):
print("hsv", i)
hsv_f = util.loading_map(lambda x: extraction.pixel_features(x, i), hsv)
validation.validate_feature(hsv_f, labels, classes, model, n_folds, False,
False, True, True)
for i in range(1, 14, 2):
print("luv", i)
luv_f = util.loading_map(lambda x: extraction.pixel_features(x, i), luv)
validation.validate_feature(luv_f, labels, classes, model, n_folds, False,
False, True, True)
#With STD
print("rgb 11 + std")
feature_extraction.calcHOGWrapper(img,
orient=8,
pixel_per_cell=5,
nr_of_cells_per_block=2)
for img in thumbsGray
]
features = HOGGray
print("Build random forest with gray features")
n_folds = 5
#iterate over multiple tree numbers 100->400
for nr_of_trees in range(100, 400, 50):
print(nr_of_trees)
model = Pipeline([("Random forest classifier",
RandomForestClassifier(n_estimators=nr_of_trees,
class_weight='balanced',
n_jobs=4))])
print("Validating Features")
validation.validate_feature(features, labels, classes, model, n_folds,
True, True, True, True)
model = Pipeline([("Extreme forest classifier",
ExtraTreesClassifier(n_estimators=nr_of_trees,
class_weight='balanced',
n_jobs=4))])
print("Validating Features")
validation.validate_feature(features, labels, classes, model, n_folds,
True, True, True, True)
hybrid_bright_hed_hog_luv = numpy.concatenate((brightness, hed_features, hog, luv_features), 1)
n_folds=5
print("Build random forest")
nr_of_trees=500
model = Pipeline([
("standard scaler", StandardScaler()),
("Random forest classifier", RandomForestClassifier(n_estimators=nr_of_trees,class_weight='balanced',n_jobs=4))
])
print("Evaluating brightness+hed+hog+luv")
validation.validate_feature(hybrid_bright_hed_hog_luv, labels, classes, model, n_folds, False, False, True, True)
print('\a')
n_folds=5
print("Build extreme forest")
nr_of_trees=500
model = Pipeline([
("standard scaler", StandardScaler()),
("Random forest classifier", ExtraTreesClassifier(n_estimators=nr_of_trees,class_weight='balanced',n_jobs=4))
])
print("Evaluating brightness+hed+hog+luv")
validation.validate_feature(hybrid_bright_hed_hog_luv, labels, classes, model, n_folds, False, False, True, True)
print('\a')
cells_per_block=(8, 8),
normalise=True), grayscaled)
print("corner features")
corners = util.loading_map(
lambda x: extraction.pixel_features(
numpy.sqrt(feature.corner_shi_tomasi(x, sigma=8)), 8), grayscaled)
n_folds = 10
model = Pipeline([("standard scaler", StandardScaler()),
("logistic regression",
LogisticRegression(solver='lbfgs',
multi_class='multinomial'))])
print("Evaluating brightness features")
validation.validate_feature(brightness, labels, classes, model, n_folds, False,
False, True, True)
print("Evaluating luv features")
validation.validate_feature(luv_features, labels, classes, model, n_folds,
False, False, True, True)
print("Evaluating hog features")
validation.validate_feature(hog, labels, classes, model, n_folds, False, False,
True, True)
print("Evaluating corner features")
validation.validate_feature(corners, labels, classes, model, n_folds, False,
False, True, True)
print("Evaluating HSV_11_std features")
validation.validate_feature(numpy.array(hsv_11_std), labels, classes, model,
n_folds, False, False, True, True)
#we combine the 2 best ones
hog_hs = numpy.concatenate((hog, hsv_11_std), 1)