def hog(self):
print('---HOG---')
img_to_hog = lambda img: utils.hog(img)
self.img_set = np.array(list(map(img_to_hog, self.img_set)))
with open('para.json', 'r') as f:
js = json.load(f)
block_wid = js['hog_dic']['block_width']
block_hei = js['hog_dic']['block_height']
n_ori = js['hog_dic']['n_orient']
self.img_set = self.img_set.reshape(
(-1, n_ori, int(self.width / block_wid + 0.5),
int(self.height / block_hei + 0.5)))
self.img_set = np.transpose(self.img_set, (0, 2, 3, 1))
self.n_example, self.width, self.height, self.color = self.img_set.shape
self.n_size_reset()
def load_img(self, path, render=True, hog=False):
try:
img = mpimg.imread(path)
if render: implot = self.mplw.canvas.imax.imshow(img, cmap="gray")
if hog:
return utils.hog(img)
else:
arr = img
if len(img.shape) == 3: arr = np.average(arr, axis=2)
if img.dtype == "float32": arr = np.floor(256*arr)
hist = np.histogram(arr.flatten(), bins=256)[0]
if render:
self.mplw.canvas.hsax.clear()
x = np.linspace(0,254,255, dtype="int")
self.mplw.canvas.hsax.bar(x, hist[:-1])
self.mplw.canvas.draw()
return hist[:-1]
except Exception as e:
print(e)
self.log(f"Invalid image file ({e})")
def hog_sliding_window(img, classifier, window_size, stride, boundary): np_img = np.array(img) img_w, img_h, _ = np_img.shape score_board = np.zeros(np_img.shape[0:2]) count_board = np.zeros(np_img.shape[0:2]) for window in window_size: w, h = window for x in range(0, img_h-h, stride): for y in range(0, img_w-w, stride): count_board = plus1(count_board, (x,y,w,h)) croped_img = utils.seperate_region(img, (x, y, w, h)) croped_img = croped_img.resize((64,64)) hog = utils.hog(np.array(croped_img)) hog = hog.reshape((-1,8,4,4)) hog = np.transpose(hog, (0,2,3,1)) if classifier(hog): score_board = plus1(score_board, (x,y,w,h)) b_img = (score_board / count_board) > boundary b_img = b_img.reshape((img_w, img_h, 1)) b_img = b_img * img result_image = Image.fromarray(b_img) return result_image
def face_detect_patch(patch, w, theta):
patch = patch / 255.0
hog_patch = hog(patch)
return prediction(np.transpose(np.concatenate(([[1]], hog_patch))), w,
theta)
from sklearn.externals import joblib as jb
from sklearn.svm import LinearSVC
import utils
import argparse
argp = argparse.ArgumentParser()
argp.add_argument("-d", "--dataset", required=True, help="dataset file")
argp.add_argument("-m",
"--model",
required=True,
help="Path model being saved")
args = vars(argp.parse_args())
Numbers, Labels = utils.getting_data(args["dataset"])
data = []
for image in Numbers:
image = utils.norm_img(image, 20)
image = utils.mass_center(image, (20, 20))
dat = utils.hog(image,
orientations=18,
pixelsPerCell=(5, 5),
cellsPerBlock=(1, 1),
normalize=True)
data.append(dat)
model = LinearSVC(random_state=40)
model.fit(data, Labels)
jb.dump(model, args["model"])