著者：飯尾 淳
本連載では「Pythonを昔から使っているものの、それほど使いこなしてはいない」という筆者が、いろいろな日常業務をPythonで処理することで、立派な「蛇使い」に育つことを目指します。その過程を温かく見守ってください。皆さんと共に勉強していきましょう。第18回では、米Google社が提供する機械学習フレームワーク「MediaPipe」の顔検出機能を用いて、お遊びのプログラム作成に挑戦します。
シェルスクリプトマガジン Vol.88は以下のリンク先でご購入できます。
図1 顔ランドマーク検出をするコード
import cv2
import mediapipe as mp
spec = mp.solutions.drawing_utils.DrawingSpec(thickness=1,
circle_radius=1)
cap = cv2.VideoCapture(0)
with mp.solutions.face_mesh.FaceMesh(
min_detection_confidence=0.5,
min_tracking_confidence=0.5) as face_mesh:
while True:
success, image = cap.read()
if not success: print("Warning: No camera image"); continue
image = cv2.cvtColor(cv2.flip(image, 1), cv2.COLOR_BGR2RGB)
image.flags.writeable = False
results = face_mesh.process(image)
# Annotate the face mesh
image.flags.writeable = True
image = cv2.cvtColor(image, cv2.COLOR_RGB2BGR)
if results.multi_face_landmarks:
for landmarks in results.multi_face_landmarks:
mp.solutions.drawing_utils.draw_landmarks(
image=image, landmark_list=landmarks,
connections=mp.solutions.face_mesh.FACEMESH_TESSELATION,
landmark_drawing_spec=spec,
connection_drawing_spec=spec)
cv2.imshow('MediaPipe FaceMesh', image)
if cv2.waitKey(5) & 0xFF == 27: break
cap.release()
図4 「顔ハメ」アプリのコード
import cv2
import numpy as np
from PIL import Image, ImageDraw
import mediapipe as mp
def scaleToHeight(img, height):
h, w = img.shape[:2]
width = round(w * (height / h))
dst = cv2.resize(img, dsize=(width, height))
return dst
def convertToRGBA(src, type):
return Image.fromarray(cv2.cvtColor(src, type))\
.convert('RGBA')
def trimOutside(base, over, loc):
drw = ImageDraw.Draw(base)
drw.rectangle([(0, 0), (loc[0]-1, over.size[1]-1)],
fill=(0, 0, 0))
drw.rectangle([(loc[0]+over.size[0], 0),
(base.size[0]-1,over.size[1]-1)],
fill=(0, 0, 0))
def overlayImage(src, overlay, location):
# convert images to PIL format
pil_src = convertToRGBA(src, cv2.COLOR_BGR2RGB)
pil_overlay = convertToRGBA(overlay, cv2.COLOR_BGRA2RGBA)
# conpose two images
pil_tmp = Image.new('RGBA', pil_src.size, (0, 0, 0, 0))
pil_tmp.paste(pil_overlay, location, pil_overlay)
trimOutside(pil_tmp, pil_overlay, location)
result_image = Image.alpha_composite(pil_src, pil_tmp)
# convert result to OpenCV format
return cv2.cvtColor(np.asarray(result_image),
cv2.COLOR_RGBA2BGRA)
def decrementTimer(timer, image, p_idx):
h, w = image.shape[:2]
if timer < 0:
p_idx = (p_idx + 1) % len(panels)
return TIMER_INITIAL_VALUE, p_idx
elif timer == 30:
global still
still = image
cv2.rectangle(image, (0, 0), (w, h), (255,255,255),
thickness=-1)
return timer - 1, p_idx
elif timer < 30:
image = still
return timer - 1, p_idx
d, r = timer // 30, timer % 30
c = 255 / 60 * r + 128
cv2.putText(image, org=(int(w/2-100), int(h/2+100)),
text=str(d),
fontFace=cv2.FONT_HERSHEY_DUPLEX,
fontScale=10.0, color=(c, c, c),
thickness=30,
lineType=cv2.LINE_AA)
return timer - 1, p_idx
# prepare the kao_hame_panels
panels = []
panels.append(cv2.imread('img1.png', cv2.IMREAD_UNCHANGED))
panels.append(cv2.imread('img2.png', cv2.IMREAD_UNCHANGED))
panels.append(cv2.imread('img3.png', cv2.IMREAD_UNCHANGED))
# capture from a camera
cap = cv2.VideoCapture(0)
ret, frame = cap.read()
# rescale the kao_hame image
height, width = frame.shape[:2]
for i in range(len(panels)):
panels[i] = scaleToHeight(panels[i], height)
p_idx = 0
panel = panels[p_idx]
p_height, p_width = panel.shape[:2]
# timing counter
TIMER_INITIAL_VALUE = 119
timer = TIMER_INITIAL_VALUE
with mp.solutions.face_mesh\
.FaceMesh(max_num_faces=1,
refine_landmarks=True,
min_detection_confidence=0.5,
min_tracking_confidence=0.5)\
as face_mesh:
while cap.isOpened():
success, image = cap.read()
if not success:
print("Ignoring empty camera frame.")
continue
image = cv2.flip(image, 1)
location = ((width-p_width)//2, 0)
image = overlayImage(image, panel, location)
image2 = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
results = face_mesh.process(image2)
if results.multi_face_landmarks != None:
timer, p_idx = decrementTimer(timer, image, p_idx)
panel = panels[p_idx]
p_height, p_width = panel.shape[:2]
else:
timer = TIMER_INITIAL_VALUE # reset timer
# triming the image
image = image[0:p_height,
location[0]:location[0]+p_width]
cv2.imshow('Virtual Face-in-Hole Cutout', image)
if cv2.waitKey(1) & 0xFF == ord('q'): break
cap.release()
cv2.destroyAllWindows()
