def get_histograms(data, means, load=True, load_file='hist_save.npy',
save=True, save_file='hist_save.npy'):
data_hists = load_data(load_file)
if len(data_hists) == 0:
print "Building histograms"
spinner = Spinner()
for i, item in enumerate(data): # (img, label, SIFT descriptors) triple
hist = np.zeros(means.shape[0])
for desc in item[0]: # for each descriptor set each descriptor has 128 values)
percent = int((float(i) / float(len(data))) * 100)
spinner.spin(percent)
dists = []
for mindex, mean in enumerate(means):
dists.append((mindex, np.linalg.norm(mean - desc)))
dists = sorted(dists, key=lambda entry: entry[1])
hist[dists[0][0]] += 1
data_hists.append(np.array(hist))
np.save(save_file, np.asarray(data_hists))
return data_hists
def get_image_data(directory, selector='sift', **kwargs):
load = kwargs.pop('load', True)
save = kwargs.pop('save', True)
load_file = kwargs.pop('load_file', selector + '_data_save.npy')
save_file = kwargs.pop('save_file', selector + '_data_save.npy')
scalex = kwargs.pop('scalex', 1.0)
scaley = kwargs.pop('scaley', 1.0)
selector = selector.lower()
hog_ppc = kwargs.pop('hog_ppc', (8, 8))
hog_cpb = kwargs.pop('hog_cpb', (3, 3))
if selector not in ['sift', 'surf', 'hog']:
print "Error: selector {} not supported".format(selector)
exit(-1)
if selector in ['sift', 'surf']:
try:
inst = kwargs.pop('inst')
except KeyError:
print """Error: SIFT/SURF given as feature selection method but no instance to
a SIFT or SURF object was given"""
exit(-1)
done = 0
cant_calc = 0
spinner = Spinner();
data = load_data(load_file) if load else []
if len(data) == 0:
print "Loading images from {}".format(directory)
dims = [Image.open(filepath[0]).size for filepath in get_filenames(directory)]
mwidth, mheight = np.mean([dim[0] for dim in dims]), np.mean([dim[1] for dim in dims])
mheight = mheight - mheight % 10
mwidth, mheight = int(mwidth), int(mheight)
size = kwargs.pop('size', (mwidth, mheight))
print "Scaling images to {} and by a factor of {}x, {}y".format(size, scalex, scaley)
print "Selecting {} features".format(selector)
if selector == 'hog':
print "pixels per cell: {}, cells per block: {}".format(hog_ppc, hog_cpb)
for filepath, label in get_filenames(directory):
spinner.spin(done, pct=False)
filename = filepath.split('/')[-1]
img = cv2.imread(filepath)
img = cv2.resize(img, size)
if scalex != 1.0 or scaley != 1.0:
img = cv2.resize(img, (0, 0), fx=scalex, fy=scaley)
if selector == 'sift' or selector == 'surf':
descriptors = get_descriptors(img, inst)
if descriptors is not None:
data.append((descriptors, label))
else:
cant_calc += 1
elif selector == 'hog':
descriptors = feature.hog(color.rgb2gray(img),
pixels_per_cell=hog_ppc, cells_per_block=hog_cpb)
data.append((descriptors, label))
elif selector == 'log':
print "LoG feature detection is not yet implemented"
exit(-1)
done += 1
if save:
print "Saving {} features to {}".format(len(data), save_file)
np.save(save_file, data)
if cant_calc > 0:
print "Unable to calculate features for {} images".format(cant_calc)
return data