目录：
1.官网的数据解释
正文：
1.官网数据解释
与FDDB数据集一样，WIDER FACE数据集的官网上也向我们提供了人脸检测图片数据集，不同的是，WIDER FACE数据集还将图片分为了训练、验证和测试数据集，不需要自己分了。不过，测试数据集是没有人脸标注的。下面是WIDER FACE数据集的标注格式。0–Parade/0_Parade_marchingband_1_849.jpg
1
449 330 122 149 0 0 0 0 0 0
与FDDB数据集一样，第一行表示图片的名称。第二行表示这张图片拥有的人脸数。因为这张图片有一张人脸，所以下面的一行是对人脸位置的标注；与FDDB不同的是，第三行的人脸标注分别是人脸左上角的x轴、左上角y轴、人脸的宽、人脸的高、人脸是否模糊的标志、人脸表情标志、光照标志、是否有效标志、遮挡标志、姿态标志。 其实真正使用到的是前面四位，这里也对后面的标志位进行解释。
| 标志 | 数字表示的程度 |
|---|---|
| 人脸是否模糊的标志 | 0表示清晰，1表示轻度模糊，2表示重度模糊 |
| 人脸表情标志 | 0表示正常表情，1表示夸张表情 |
| 光照标志 | 0表示正常光照，1表示严重光照 |
| 是否有效标志 | 0表示有效，1表示无效 |
| 遮挡标志 | 0表示没有遮挡，1表示部分遮挡，2表示严重遮挡 |
| 姿态标志 | 0表示正常姿态，1表示非正常姿态 |
2.处理数据
与FDDB数据类似，人脸的获取也用矩形框进行裁剪。以x、y、w、h代表前四位标志，获取方式如下。x1 = x
y1 = y
x2 = x + w
y2 = y + h
face_image = image[y1:y2,x1:x2] #注意x y的顺序
在FDDB数据集处理时（FDDB数据集处理）已经给出了获取FDDB非人脸的程序框图，它也适用与WIDER FACE数据集。下面给出如何处理标注人脸的txt文件的程序框图，FDDB数据集的处理方法也是一致的。
3.处理好的数据分享
下面是我处理得到的人脸以及非人脸训练和验证数据。有需要的可以下载。WIDER FACE 人脸/非人脸数据
密码：b7qcy9
4.代码
import os
import cv2
from datetime import datetime
import numpy as np
print(datetime.now())
def get_image_name(image_path):
"""获取图片名称"""
strat_index = 0
end_index = 0
for i in range(len(image_path)):
if image_path[i] == '/':
strat_index = i
if image_path[i] == '.':
end_index = i
image_name = image_path[strat_index+1: end_index]
return image_name
#测试截取所有人脸-------------------------------------------------------------
def get_random(image_w, image_h, bounding_w, bounding_h):
"""产生随机数"""
#图像的宽 图像的高 框的宽 框的高
limx = image_w - bounding_w
limy = image_h - bounding_h
x1 = np.random.randint(0, limx)
x2 = x1 + bounding_w
y1 = np.random.randint(0, limy)
y2 = y1 + bounding_h
return x1, y1, x2, y2
def compuet_iou(rect1, rect2):
"""计算两个矩形的IOU"""
#rect是矩形的四个坐标 左上角和右下角坐标 rect = [(x,1,y2),(x2,y2)]
area1 = (rect1[3] - rect1[1]) * (rect1[2] - rect1[0])
area2 = (rect2[3] - rect2[1]) * (rect2[2] - rect2[0])
all_area = area1 + area2
x1 = max(rect1[0], rect2[0])
y1 = max(rect1[1], rect2[1])
x2 = min(rect1[2], rect2[2])
y2 = min(rect1[3], rect2[3])
h = max(0, y2 - y1)
w = max(0, x2 - x1)
overlap = h * w
try :
iou = overlap/(all_area - overlap)
except:
iou = 1
else :
iou = overlap/(all_area - overlap)
return iou
# 获取所有文件名地址、文件名下标、人脸数、人脸数下标
def get_img(label_text):
image_path = [] #存储地址
image_path_index = [] #存储txt地址所在行
face_num = [] #存储人脸数
row_num =0
with open(label_text) as obj:
for l in obj.readlines():
row_num += 1
if '.' in l: #图片地址
image_path.append(l.strip('\n')) #去掉换行符
image_path_index.append(row_num)#地址所在行数
flag = True #这个标志位使得只读取图片地址的下一行
else:
if flag:
face_num.append(int(l))
flag =False
face_num_index = [x + 1 for x in image_path_index] #人脸数所在行数
return image_path, image_path_index,face_num,face_num_index
label_text = 'wider_face_train_bbx_gt.txt'
image_path, image_path_index,face_num,face_num_index = get_img(label_text)
# ten_face = []
for i in range(len(face_num)):
if face_num[i] == 1:
print(image_path[i])
if i ==2:
break
# ten_face.append(i)
# print(ten_face) #[279, 3808, 7512, 9227] 没有人脸
#提取框坐标（三维数组）---------------------------------------------------------
def get_bouning(label_text):
image_path, image_path_index,face_num,face_num_index = get_img(label_text)
mid_bounding_box = []
bounding_box = []
true_bounding = [] #小于10的人脸框
for i in range(len(face_num_index)):
strat_index = face_num_index[i]
end_index = face_num_index[i] + face_num[i]
# print(strat_index,end_index)
with open(label_text) as obj:
for l in obj.readlines()[strat_index: end_index]:
#读取框数据
begin_bounding_index = 0
for i in range(len(str(l))) : #遍历某一行
if l[i] == ' ' :
end_bounding_index = i
mid_bounding_box.append(int(
str(l)[begin_bounding_index: end_bounding_index]))
begin_bounding_index = end_bounding_index
bounding_box.append(mid_bounding_box[:4])
mid_bounding_box = [] #清空
true_bounding.append(bounding_box)
bounding_box = [] #清空
return true_bounding,image_path, image_path_index,face_num,face_num_index
# print(len((true_bounding)[2]))
# print(image_path)
#从图片中框出人脸（截取）-------------------------------------------------------
def get_face(label_text,save_face_path,root_path):
true_bounding,image_path, image_path_index,face_num,face_num_index =\
get_bouning(label_text)
length = len(image_path)
for i in range(length):
if face_num[i] < 5 and face_num[i] > 0: #为了防止脸数太多太小 剔除大于5张脸的
face_rect = []
all_path = os.path.join(root_path, image_path[i]) #所有路径
image_raw_name = get_image_name(image_path[i])
# save_image_name
image = cv2.imread(all_path, 1)
image_h, image_w , _ = image.shape
for j in range(len(true_bounding[i])) : #多个人脸
x = int(true_bounding[i][j][0])
y = int(true_bounding[i][j][1])
w = int(true_bounding[i][j][2])
h = int(true_bounding[i][j][3])
face_flag = 'face_' + str(j) + '.jpg'
image_name = image_raw_name + face_flag #1
save_image_name = os.path.join(save_face_path, image_name)#1
# print(save_image_name)
#框出
# image = cv2.rectangle(image, (x, y), (x+w,y+h), (0,255,0), 2)
image_cut = image[y: y+h, x: x+w]
rect = (x,y,x+w,y+h)
face_rect.append(rect)
if h>100 or w>100:
try :
cv2.imwrite(save_image_name, image_cut)
except:
print(save_image_name)
else:
cv2.imwrite(save_image_name, image_cut)
def get_no_face(label_text,save_no_face_path,root_path):
true_bounding,image_path, image_path_index,face_num,face_num_index =\
get_bouning(label_text)
length = len(image_path)
for i in range(length):
if face_num[i] < 5 and face_num[i] > 0: #为了防止脸数太多太小 剔除大于5张脸的
face_rect = []
all_path = os.path.join(root_path, image_path[i]) #所有路径
image_raw_name = get_image_name(image_path[i])
image = cv2.imread(all_path, 1)
image_h, image_w , _ = image.shape
for j in range(len(true_bounding[i])) : #多个人脸
x = int(true_bounding[i][j][0])
y = int(true_bounding[i][j][1])
w = int(true_bounding[i][j][2])
h = int(true_bounding[i][j][3])
rect = (x,y,x+w,y+h)
face_rect.append(rect)
# 截取非人脸
flag = True
no_face_number = 0
lim_while = 0 #循环10次没有就下次个图片
while flag:
lim_while += 1
if lim_while == 10:
break
# 增加判断 防止随机数产生失败
ious = []
try:
no_x1, no_y1, no_x2, no_y2 = get_random(image_w, image_h, w, h)
except:
break
else:
no_x1, no_y1, no_x2, no_y2 = get_random(image_w, image_h, w, h)
rect2 = (no_x1, no_y1, no_x2, no_y2) #获取随机矩形框
for rect in face_rect:
iou = compuet_iou(rect, rect2)
ious.append(iou)
iou = max(ious) #取最大的iou
if iou<0.2:
no_face_number += 1
no_face_image = image[no_y1: no_y2, no_x1: no_x2]
no_face_flag = 'no_face_' + str(no_face_number) + '.jpg'
image_name2 = image_raw_name + no_face_flag#0
save_image_name2 = os.path.join(save_no_face_path, image_name2)#0
cv2.imwrite(save_image_name2, no_face_image)
if no_face_number == len(true_bounding[i]) :#1:3 face:no_face 设置产生非人脸个数
flag = False
#主程序
#训练数据集
# label_text = 'wider_face_train_bbx_gt.txt'
# root_path = r"C:\Users\user\Desktop\face_data\WIDER\WIDER_train\images"
# save_path = "pr_face"
# save_path2 = "pr_no_face"
# start_time = datetime.now()
# get_face(label_text,save_path, root_path)
# end_time = datetime.now()
# use_time = str(end_time - start_time)
# print('截取并保存人脸图片所用时间时间为:' + use_time)
# start_time = datetime.now()
# get_no_face(label_text,save_path2,root_path)
# end_time = datetime.now()
# use_time = str(end_time - start_time)
# print('截取并保存非人脸图片所用时间时间为:' + use_time)
# 验证数据集
# val_label_text = 'wider_face_val_bbx_gt.txt'
# val_root_path = r"C:\Users\user\Desktop\face_data\WIDER\WIDER_val\images"
# vaL_save_path = "pr_face"
# val_save_path2 = "pr_no_face"
# start_time = datetime.now()
# get_face(val_label_text,vaL_save_path, val_root_path)
# end_time = datetime.now()
# use_time = str(end_time - start_time)
# print('截取并保存验证人脸图片所用时间时间为:' + use_time)
# start_time = datetime.now()
# get_no_face(val_label_text,val_save_path2,val_root_path)
# end_time = datetime.now()
# use_time = str(end_time - start_time)
# print('截取并保存验证非人脸图片所用时间时间为:' + use_time)
# # 测试数据集
# test_label_text = 'wider_face_test_filelist.txt'
# test_root_path = r"C:\Users\user\Desktop\face_data\WIDER\WIDER_test\images"
# test_save_path = "wider_test_face_data"
# test_save_path2 = "wider_test_no_face_data"
# start_time = datetime.now()
# get_face(test_label_text,test_save_path, test_save_path2)
# end_time = datetime.now()
# use_time = str(end_time - start_time)
# print('截取并测试保存人脸图片所用时间时间为:' + use_time)
# start_time = datetime.now()
# get_no_face(test_label_text,test_save_path2,test_save_path2)
# end_time = datetime.now()
# use_time = str(end_time - start_time)
# print('截取并保存测试非人脸图片所用时间时间为:' + use_time)
# 查看文件夹人脸数
# count =0
# for root,dirs,files in os.walk('pr_no_face'):
# for each in files:
# count += 1
# print ('得到人脸图片个数为:' + str(count) ) #7784 1892
# count =0
# for root,dirs,files in os.walk('wider_val_no_face_data'):
# for each in files:
# count += 1
# print ('得到非人脸图片个数为:' + str(count) ) # 6562
结语：
如果对你有帮助，就给我一个赞吧，如何有问题，可以在评论区进行讨论。
上一篇：[基于tensorflow的人脸检测] 基于神经网络的人脸检测2——FDDB数据集处理
下一篇：[基于tensorflow的人脸检测] 基于神经网络的人脸检测4——数据集的标签生成
扫一扫
5434
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
