from numpy import mean, std
from sklearn.model_selection import cross_val_score
from classification.pipe import get_pipe
from datagen.DatasetGenerator import DatasetGenerator
from datagen.FileProviderAsl import FileProviderAsl
from preprocessing.PreProcessorAsl import PreProcessorAsl
from preprocessing.representation.BagOfHogs import BagOfHogs
g = DatasetGenerator(
FileProviderAsl("../../../resource/dataset/fingerspelling5/dataset5/"),
preprocessor=PreProcessorAsl(descriptor=BagOfHogs(
"../../../resource/models/codebook_total_new.pkl")))
data, labels = g.generate()
model = get_pipe()
results = cross_val_score(model, data, labels.ravel(), cv=6)
print("Accuracy: " + str(mean(results)) + "(+/- " + str(std(results)))
def colourbased_skin_segmentation(image,
lower_thresh=np.array([0, 80, 80],
dtype="uint8"),
upper_thresh=np.array([20, 240, 240],
dtype="uint8")):
converted = cv2.cvtColor(image, cv2.COLOR_BGR2HSV)
skin_mask = cv2.inRange(converted, lower_thresh, upper_thresh)
skin_mask = cv2.GaussianBlur(skin_mask, (3, 3), 2)
# apply a series of erosions and dilations to the mask
# using an elliptical kernel
kernel = cv2.getStructuringElement(cv2.MORPH_ELLIPSE, (3, 3))
skin_mask = cv2.erode(skin_mask, kernel, iterations=1)
skin_mask = cv2.dilate(skin_mask, kernel, iterations=1)
# blur the mask to help remove noise, then apply the
# mask to the frame
_, likelihood = cv2.threshold(skin_mask, 10, 255, cv2.NORM_MINMAX)
return likelihood
example_image_file = "../../../resource/examples/color_0_0028.png"
# read image
img, _ = FileProviderAsl.read_img(example_image_file)
cv2.imshow('image', img)
img = colourbased_skin_segmentation(img)
cv2.imshow("Likelihood", img.astype(np.uint8))
cv2.waitKey(0)
import random
from datagen.FileProviderAsl import FileProviderAsl
from preprocessing.PreProcessorAsl import PreProcessorAsl
from preprocessing.representation.BagOfHogs import BagOfHogs
from preprocessing.segmentation.MRFAsl import MRFAsl
codebook_file = '../../resource/models/codebook.pkl'
paths = FileProviderAsl(
"../../resource/dataset/fingerspelling5/dataset5/").img_file_paths
paths = random.sample(paths[random.choice(list(paths.keys()))], k=100)
img = FileProviderAsl.read_img(random.choice(paths))
pp = PreProcessorAsl(segmenter=MRFAsl(), descriptor=BagOfHogs(codebook_file))
img = pp.preprocess(img)
descriptor = pp.get_descr(img)
print(str(descriptor))
from sklearn.externals import joblib
from datagen.FileProviderAsl import FileProviderAsl
from preprocessing.PreProcessorAsl import PreProcessorAsl
from preprocessing.representation.BagOfHogs import BagOfHogs
paths_dict = FileProviderAsl(
"../../resource/dataset/fingerspelling5/dataset5/").img_file_paths
codebook_file = '../../resource/models/codebook_total.pkl'
images = []
for letter in paths_dict:
for path in paths_dict[letter]:
images.append(FileProviderAsl.read_img(path))
break
codebook = BagOfHogs(n_bins=16, winsize=6).get_codebook(
PreProcessorAsl(img_size=(60, 60)).preprocess_all(images), k_words=512)
joblib.dump(codebook, codebook_file)
import cv2
import numpy as np
from datagen.FileProviderAsl import FileProviderAsl
from preprocessing.segmentation.MRFAsl import MRFAsl
example_image_file = "../../../resource/examples/color_0_0028.png"
# read image
img, img_depth = FileProviderAsl.read_img(example_image_file)
cv2.imshow('image', img)
img_segment = cv2.GaussianBlur(img, (11, 11), 2)
threshold = img_depth[int(img_depth.shape[0] / 2), int(img_depth.shape[1] / 2)]
img_gray = cv2.cvtColor(img_segment, cv2.COLOR_RGB2GRAY)
background_label = img_gray.copy()
background_label[img_depth != threshold] = 0
background_label[img_gray > 235] = 0
background_label[img_gray < 25] = 0
area_foreground = background_label.copy()
area_foreground[0:int(img_segment.shape[0] / 2 - 30),
0:int(img_segment.shape[1] / 2 - 30)] = 0
area_foreground[int(img_segment.shape[0] / 2) + 30:img_segment.shape[0],
int(img_segment.shape[1] / 2) + 30:img_segment.shape[0]] = 0
pixels_fg = img_segment[area_foreground != 0].reshape(-1, 3)
pixels_bg = img_segment[background_label == 0].reshape(-1, 3)
import cv2
import numpy as np
from numpy.linalg import norm
from numpy.ma import floor
from datagen.FileProviderAsl import FileProviderAsl
from preprocessing.segmentation.ColourHistogram import ColourHistogram
paths_dict = FileProviderAsl(sets=["E"], alphabet=["a"]).img_file_paths
winsize = 3
samples = []
for paths in paths_dict.values():
for path in paths:
img = read_image(path)
img_filtered = cv2.pyrMeanShiftFiltering(img, 5, 50)
y = int(floor(img.shape[0] / 2) - floor(winsize / 2))
x = int(floor(img.shape[1] / 2) - floor(winsize / 2))
# sample = img.copy()
# cv2.rectangle(sample, (x, y), (x+winsize, y+winsize), (0, 255, 0))
# cv2.imshow("Box", sample)
window = img[y:y + winsize, x:x + winsize]
samples.append(window.reshape(-1, 3))
n_bins = 32
bin_range = 255 / n_bins
hist = np.zeros(shape=(1, 3 * n_bins))
for sample in samples:
hist += ColourHistogram.colour_hist(sample)
from datagen.DatasetGenerator import DatasetGenerator
from datagen.FileProviderAsl import FileProviderAsl
from preprocessing.PreProcessorAsl import PreProcessorAsl
from preprocessing.representation.HistogramOfGradients import HistogramOfGradients
from preprocessing.segmentation.MRFAsl import MRFAsl
g = DatasetGenerator(file_provider=FileProviderAsl(
"../../resource/dataset/fingerspelling5/dataset5/"),
preprocessor=PreProcessorAsl(
segmenter=MRFAsl(),
descriptor=HistogramOfGradients(window_size=6,
n_bins=8)))
data, labels = g.generate(sample_size=2500)
g.save(data, labels, '../../resource/models/', 'hog')