Skip to content

Instantly share code, notes, and snippets.

Embed
What would you like to do?
Detect face in image and save cropped.
#! /usr/bin/env python
"""
Detect face and save cropped images.
http://vaaaaaanquish.hatenablog.com/entry/2016/08/15/193636
http://dlib.net/face_detector.py.html
"""
import cv2
import os
import glob
import math
import matplotlib.pyplot as plt
from matplotlib import patches
import dlib
# % matplotlib inline
def add_bboxes(ax, bbox, rect_opt=None, figsize=(10, 10)):
# Add bounding boxes
rect_opt = rect_opt or {}
xy = bbox[:2]
width, height = bbox[2:]
opt = {'linewidth': 2, 'edgecolor': 'r',
'facecolor': 'none'}
opt.update(rect_opt)
bbox = patches.Rectangle(xy, width, height, **opt)
ax.add_patch(bbox)
return ax
def get_text_pos(img_w, img_h, bbox, loc='in'):
left, top, width, height = bbox
if loc == 'right':
hrz_offset = img_w / 10
vtc_offset = height / 2
x = left + width + hrz_offset
y = top + vtc_offset
opt = {}
elif loc == 'in':
hrz_offset = width / 2
vtc_offset = height / 2
x = left + hrz_offset
y = top + vtc_offset
opt = {'horizontalalignment': 'center'}
return x, y, opt
def adjust_bbox(bbox):
bbox_new = list(bbox)
w_ratio = 0.1
h_ratio = 0.2
bbox_new[0] -= bbox_new[2] * w_ratio
bbox_new[2] += bbox_new[2] * w_ratio * 2
bbox_new[1] -= bbox_new[3] * h_ratio
bbox_new[3] += bbox_new[3] * h_ratio * 2
return bbox_new
def add_facedetected_rect(ax, img_size,
dets, scores, indices, show_info=True, text_opt_add=None):
img_h, img_w = img_size
for rect, score, idx in zip(dets, scores, indices):
bbox = rect2bbox(rect)
ax = add_bboxes(ax, bbox)
if show_info:
text = 'idx: {}\nscore: {:.3f}'.format(idx, score)
x, y, opt = get_text_pos(img_h, img_w, bbox)
text_opt = {'color': 'b', 'fontsize': 10}
text_opt.update(opt)
text_opt_add = {} if text_opt_add is None else text_opt_add
text_opt.update(text_opt_add)
ax.text(x, y, text, **text_opt)
return ax
def check_img(img):
assert img is not None
assert img.ndim == 3
def imread_rgb(path_img):
img = cv2.imread(path_img)
if img is None:
raise IOError
else:
return cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
def plot_detected_face(path_images, detector, save_fig=None, show_info=True):
n = len(path_images)
ncols = 5
nrows = math.ceil(n / ncols)
figsize = (3 * ncols, 3 * nrows)
fig, axes = plt.subplots(nrows=nrows, ncols=ncols, figsize=figsize)
for ax, path_img in zip(axes.ravel(), path_images):
try:
img = imread_rgb(path_img)
check_img(img)
opt = (1, 0) # thresh
dets, scores, indices = detector.run(img, *opt)
ax.imshow(img)
add_facedetected_rect(ax, img.shape[:2], dets, scores, indices,
show_info)
ax.axis('off')
except (IOError, AssertionError):
errmsg = 'Broken File: {}'.format(path_img)
print(errmsg)
pass
plt.subplots_adjust(wspace=0, hspace=0)
if save_fig:
plt.savefig(save_fig)
plt.show()
def bbox2start_end(bbox):
left, top, width, height = bbox
x_start = left
x_end = left + width
y_start = top
y_end = top + height
return x_start, x_end, y_start, y_end
def get_img_cropped(img, bbox):
x_start, x_end, y_start, y_end =\
bbox2start_end(bbox)
return img[y_start:y_end, x_start:x_end]
def rect2bbox(rect):
width = rect.right() - rect.left()
height = rect.bottom() - rect.top()
return rect.left(), rect.top(), width, height
def save_detected_face(path_images, detector, out_dir,
min_h=256, min_w=256, verbose=True):
n_imgs_saved = 0
for path_img in path_images:
try:
img = imread_rgb(path_img)
check_img(img)
opt = (1, 0) # thresh
dets, scores, indices = detector.run(img, *opt)
if dets:
filename = os.path.basename(path_img)
file_parts = os.path.splitext(filename)
for i, rect in enumerate(dets):
bbox = rect2bbox(rect)
if bbox[2] >= min_w and bbox[3] >= min_h:
img_cropped = get_img_cropped(img, bbox)
save_path = os.path.join(out_dir,
'{}_{}{}'.format(file_parts[0], i, file_parts[1]))
cv2.imwrite(save_path,
cv2.cvtColor(img_cropped, cv2.COLOR_RGB2BGR))
n_imgs_saved += 1
if verbose:
print('Images #{} saved: {}'.format(n_imgs_saved, save_path))
except (IOError, AssertionError):
errmsg = 'Broken File: {}'.format(path_img)
print(errmsg)
pass
def main():
IMG_EXT = ('.JPG', '.jpg', '.jpeg', '.png', '.PNG')
keys = ('glasses', 'man', 'woman', 'no_glass', 'sunglass', 'man_itoh',
'woman_itoh', 'glasses_itoh')
for key in keys:
dir_in = 'data/{}'.format(key)
dir_out = 'data/{}_cropped'.format(key)
print("-" * 40)
print('key: {} START'.format(key))
print("-" * 40)
os.makedirs(dir_out, exist_ok=True)
# Load images
path_images = [os.path.join(dir_in, file_path)
for file_path in os.listdir(dir_in)
if os.path.isfile(os.path.join(dir_in, file_path)) and \
os.path.splitext(file_path)[-1] in IMG_EXT]
# Build detector object
detector = dlib.get_frontal_face_detector()
# Check how detected
# plot_detected_face(path_images[:20], detector)
# Save cropped images
save_detected_face(path_images, detector, dir_out)
if __name__ == '__main__':
main()
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
You can’t perform that action at this time.