人臉區域添加矩形黑框
- 依賴：Dlib library
- Dlib簡介：Dlib is a modern C++ toolkit containing machine learning algorithms and tools for creating complex software in C++ to solve real world problems. See http://dlib.net for the main project documentation and API reference.
- 基本思路：先使用dlib檢測出人臉關鍵點，根據關鍵點的位置，確定框的中心點以及寬高；
- 主要mask區域有：左眼，右眼，鼻子，嘴部，臉頰
#coding=utf-8
import os
import sys
import cv2
import dlib
import numpy as np
# change directory to this module path
try:
this_file = __file__
except NameError:
this_file = sys.argv[0]
this_file = os.path.abspath(this_file)
#print this_file
if os.path.dirname(this_file):
os.chdir(os.path.dirname(this_file))
script_dir = os.getcwd()
#print script_dir
def get_rect_mask(shape,index_area=0,h = None,w = None,w_ratio = None,h_ratio = None):
"""
shape : image of landmarks/parts
index_area:
0:refer to left eye
1:refer to right eye
2:refer to nose;
3:refer to mouth or face;
"""
if index_area==0:#36-41 six points
if w_ratio is None or h_ratio is None:
w_ratio = np.random.choice([2.0, 2.05, 2.10,2.15],1,p =[0.2,0.4,0.2,0.2])[0]
h_ratio = np.random.choice([2.3, 2.35, 2.25,2.38],1,p =[0.2, 0.4, 0.25, 0.15])[0]
x_center = int(shape.part(36).x + shape.part(37).x + shape.part(39).x)/3
y_center = int(shape.part(37).y+shape.part(41).y)/2
width = int(w_ratio*(shape.part(39).x-shape.part(36).x)/2)
height = int(h_ratio*(shape.part(41).y-shape.part(37).y)/2)
x0 = max(x_center - width,0)
y0 = max(y_center - height,0)
x1 = min(x_center + width,w)
y1 = max(y_center - height,0)
x2 = min(x_center + width,w)
y2 = min(y_center + height,h)
x3 = max(x_center - width,0)
y3 = min(y_center + height,h)
pts = np.asarray([[x0,y0],[x1,y1],[x2,y2],[x3,y3]])
return pts
elif index_area==1:#42-47 six points
if w_ratio is None or h_ratio is None:
w_ratio = np.random.choice([2.0, 2.05, 2.10,2.15],1,p =[0.2,0.4,0.2,0.2])[0]
h_ratio = np.random.choice([2.3, 2.35, 2.25, 2.38],1,p =[0.2, 0.4, 0.25, 0.15])[0]
x_center = int(shape.part(42).x + shape.part(45).x + shape.part(44).x)/3
y_center = int(shape.part(43).y+shape.part(47).y)/2
width = int(w_ratio*(shape.part(45).x-shape.part(42).x)/2)
height = int(h_ratio*(shape.part(47).y-shape.part(43).y)/2)
x0 = max(x_center - width,0)
y0 = max(y_center - height,0)
x1 = min(x_center + width,w)
y1 = max(y_center - height,0)
x2 = min(x_center + width,w)
y2 = min(y_center + height,h)
x3 = max(x_center - width,0)
y3 = min(y_center + height,h)
pts = np.asarray([[x0,y0],[x1,y1],[x2,y2],[x3,y3]])
return pts
elif index_area==2:#27-35 nine points
if w_ratio is None or h_ratio is None:
w_ratio = np.random.choice([1.3,1.4,1.45],1,p=[0.3,0.4,0.3])[0]
h_ratio = np.random.choice([0.8,0.9,1.0],1,p=[0.5,0.4,0.1])[0]
x_center = int(shape.part(31).x+shape.part(35).x+shape.part(27).x)/3
width = int(w_ratio*(shape.part(35).x-shape.part(31).x)/2)
y_center = int(shape.part(27).y + shape.part(29).y + shape.part(33).y)/3
height = int(h_ratio*(shape.part(33).y - shape.part(27).y+10)/2)
x0 = x_center
y0 = max(y_center - height,0)
x1 = max(x_center - width,0)
y1 = min(y_center + height,h)
x2 = min(x_center + width,w)
y2 = min(y_center + height,h)
pts = np.asarray([[x0,y0],[x1,y1],[x2,y2]])
return pts
elif index_area==3: #1.2,1.5
if w_ratio is None or h_ratio is None:
w_ratio = np.random.choice([0.95, 1.15, 1.2, 1.25],1,p=[0.25,0.2,0.25,0.3])[0]
h_ratio = np.random.choice([0.85,0.95,1.05,1.1],1,p=[0.3,0.25,0.35,0.1])[0]
x_center = int(shape.part(48).x + shape.part(54).x)/2
width = int(w_ratio*(shape.part(54).x - shape.part(48).x)/2)
y_center = int(shape.part(58).y + shape.part(50).y)/2
height = int(h_ratio*(shape.part(58).y - shape.part(50).y)/2)
x0 = max(x_center - width,0)
y0 = max(y_center - height,0)
x1 = min(x_center + width,w)
y1 = max(y_center - height,0)
x2 = min(x_center + width,w)
y2 = min(y_center + height,h)
x3 = max(x_center - width,0)
y3 = min(y_center + height,h)
pts = np.asarray([[x0,y0],[x1,y1],[x2,y2],[x3,y3]])
return pts
elif index_area==4:
if w_ratio is None or h_ratio is None:
w_ratio = np.random.choice([1.2,1.35,1.25,1.3,1.32],1,p=[0.2,0.2,0.2,0.3,0.1])[0]
h_ratio = np.random.choice([0.6,0.66,0.56,0.48,0.5],1,p=[0.2,0.2,0.3,0.2,0.1])[0]
Flags = np.random.choice([0,1],1)
#已關鍵點確定mask的中心點座標 (x_center, y_center_)
if Flags==0:
x_point = np.random.choice([3,4,5,6],2,p=[0.3,0.3,0.2,0.2])
else:
x_point = np.random.choice([10,11,12,13],2,p=[0.25,0.3,0.25,0.2])
x_center = int((shape.part(x_point[0]).x+shape.part(x_point[1]).x)/2)
y_point = np.random.choice([28,29,30,33],2,p=[0.2,0.3,0.3,0.2])
y_center = int((shape.part(y_point[0]).y + shape.part(y_point[1]).y)/2)
#中心點的左右寬度
width = int(w_ratio*(shape.part(29).x - shape.part(2).x)/2)
#中心點的上下高度
height = int(h_ratio*(shape.part(8).y - shape.part(20).y))
x0 = max(x_center - width,int(shape.part(2).x))
y0 = max(y_center - height,5)
x1 = min(x_center + width,int(shape.part(16).x))
y1 = max(y_center - height,5)
x2 = min(x_center + width,int(shape.part(16).x))
y2 = min(y_center + height,h-5)
x3 = max(x_center - width,int(shape.part(2).x))
y3 = min(y_center + height,h-5)
pts = np.asarray([[x0,y0],[x1,y1],[x2,y2],[x3,y3]])
return pts
else:
raise ValueError("parameter index_area value is invalid")
def add_mask(dir_path):
img_list = os.listdir(dir_path)
save_dir = "./occ_mask"
for item in img_list:
full_path = os.path.join(dir_path,item)
img = cv2.imread(full_path)
height,width = img.shape[:2]
print(height,width)
gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
#使用默認人臉識別的模型
detector = dlib.get_frontal_face_detector()
# 獲取人臉關鍵點預訓練模型
# 模型獲取地址：http://dlib.net/files/
predictor = dlib.shape_predictor("./config/shape_predictor_68_face_landmarks.dat")
# Ask the detector to find the bounding boxes of each face. The 1 in the
# second argument indicates that we should upsample the image 1 time. This
# will make everything bigger and allow us to detect more faces.
dets = detector(gray, 1)
#print("Number of faces detected: {}".format(len(dets)))
for k, d in enumerate(dets):
print("Left: {} Top: {} Right: {} Bottom: {}".format(
d.left(), d.top(), d.right(), d.bottom()))
# Get the landmarks/parts for the face in box d.
shape = predictor(img, d)
#print shape.part(0).x, shape.part(0).y
pts_list = []
#隨機選擇黑框的個數
nums_mask = np.random.choice([1,2,3],1,p=[0.35,0.5,0.15])[0]
#隨機選擇關鍵點的大致位置
index_list = np.random.randint(0,4,nums_mask)
#print set(index_list)
#index_list = [4]
for index in set(index_list):
#print index
pts_list.append(get_rect_mask(shape,index,height,width))
#image = cv2.fillConvexPoly(img,pts,(0,0,0))
image = cv2.fillPoly(img,pts_list,(0,0,0))
save_name = os.path.join(save_dir,item)
cv2.imwrite(save_name,image)
if __name__=="__main__":
dir_path = sys.argv[1]
add_mask(dir_path)
- 結果展示：
- Dlib關鍵點的對應關係
![在這裏插入圖片描述]()
