CK+数据集介绍
?CK+数据库是在 Cohn-Kanade Dataset 的基础上扩展来的，包含表情的label和Action Units 的label。这个数据库包括123个subjects, 593 个 image sequence，每个image sequence的最后一张 Frame 都有action units 的label，而在这593个image sequence中，有327个sequence 有 emotion的 label。这个数据库是人脸表情识别中比较流行的一个数据库，很多文章都会用到这个数据做测试。
?CK+数据库一共有4个压缩文件：
- extended-cohn-kanade-images.zip：共有123个受试者的593个序列，在峰值帧进行FACS编码。所有序列都是从中性面到峰值表情。
- Landmarks.zip：所有序列都是AAM跟踪的，每个图像有68个点的Landmarks。
- FACS_labels.zip：对于每个序列（593），只有1个FACS文件，这是最后一帧（峰值帧）。文件的每一行对应一个特定的AU，然后是强度。
- Emotion_labels.zip：593个序列中只有327个具有情感序列，情绪类别分别是0=neutral, 1=anger, 2=contempt, 3=disgust, 4=fear, 5=happy, 6=sadness, 7=surprise。
百度云链接：https://pan.baidu.com/s/182ZigVgfhmO-YnLy3ip1dQ
百度云密码：CKCK
CK+数据集存为.h5文件
?首先对CK+数据集切割后的图片文件保存为为.h5文件，具体代码如下所示：
#create data and label for CK+
#0=anger,1=disgust,2=fear,3=happy,4=sadness,5=surprise,6=contempt
#contain 135,177,75,207,84,249,54 images
import csv
import os
import numpy as np
import h5py
import skimage.io
ck_path = r'Dataset\CK+'
anger_path = os.path.join(ck_path, 'anger')
disgust_path = os.path.join(ck_path, 'disgust')
fear_path = os.path.join(ck_path, 'fear')
happy_path = os.path.join(ck_path, 'happy')
sadness_path = os.path.join(ck_path, 'sadness')
surprise_path = os.path.join(ck_path, 'surprise')
contempt_path = os.path.join(ck_path, 'contempt')
## Create the list to store the data and label information
data_x = []
data_y = []
datapath = os.path.join('H5File','CK+.h5')
if not os.path.exists(os.path.dirname(datapath)):
os.makedirs(os.path.dirname(datapath))
## order the file, so the training set will not contain the test set (don't random)
files = os.listdir(anger_path)
files.sort()
for filename in files:
I = skimage.io.imread(os.path.join(anger_path,filename))
data_x.append(I.tolist())
data_y.append(0)
files = os.listdir(disgust_path)
files.sort()
for filename in files:
I = skimage.io.imread(os.path.join(disgust_path,filename))
data_x.append(I.tolist())
data_y.append(1)
files = os.listdir(fear_path)
files.sort()
for filename in files:
I = skimage.io.imread(os.path.join(fear_path,filename))
data_x.append(I.tolist())
data_y.append(2)
files = os.listdir(happy_path)
files.sort()
for filename in files:
I = skimage.io.imread(os.path.join(happy_path,filename))
data_x.append(I.tolist())
data_y.append(3)
files = os.listdir(sadness_path)
files.sort()
for filename in files:
I = skimage.io.imread(os.path.join(sadness_path,filename))
data_x.append(I.tolist())
data_y.append(4)
files = os.listdir(surprise_path)
files.sort()
for filename in files:
I = skimage.io.imread(os.path.join(surprise_path,filename))
data_x.append(I.tolist())
data_y.append(5)
files = os.listdir(contempt_path)
files.sort()
for filename in files:
I = skimage.io.imread(os.path.join(contempt_path,filename))
data_x.append(I.tolist())
data_y.append(6)
print(np.shape(data_x))
print(np.shape(data_y))
datafile = h5py.File(datapath,'w')
datafile.create_dataset("data_pixel", dtype='uint8', data = data_x)
datafile.create_dataset("data_label", dtype='int64', data = data_y)
datafile.close()
print('Save data finish!!!')
数据存储为torch.utils.data类型
?将数据集转化为torch.utils.data类型的数据，具体的代码如下所示：
from __future__ import print_function
from PIL import Image
import numpy as np
import h5py
import torch.utils.data as data
class CK(data.Dataset):
'''
Args:
train (bool, optional): If True, creates dataset from training set, otherwise
creates from test set.
transform (callable, optional): A function/transforms that takes in an PIL image
and returns a transformed version.
there are 135, 177, 75, 207, 84, 249, 54 images in data;
we choose 123, 159, 66, 186, 75, 225, 48 images for training;
we choose 12, 8, 9, 21, 9, 24, 6 images for testing;
the split are in order according to the fold number.
'''
def __init__(self, split='Training', fold=1, transform=None):
self.transform = transform
self.split = split # training set or test set
self.fold = fold # the k-fold cross validation
self.data = h5py.File('./H5File/CK+.h5','r',driver='core')
number = len(self.data['data_label']) # 981
sum_number = [0, 135, 312, 387, 594, 678, 927, 981] # the sum of class number
test_number = [12, 18, 9, 21, 9, 24, 6] # the number of each class
test_index = []
train_index = []
for j in range(len(test_number)):
for k in range(test_number[j]):
if self.fold != 10: #the last fold start from the last element
test_index.append(sum_number[j]+(self.fold-1)*test_number[j]+k)
else:
test_index.append(sum_number[j+1]-1-k)
for i in range(number):
if i not in test_index:
train_index.append(i)
#now load the picked numpy arrays
if self.split == 'Training':
self.train_data = []
self.train_labels = []
for ind in range(len(train_index)):
self.train_data.append(self.data['data_pixel'][train_index[ind]])
self.train_labels.append(self.data['data_label'][train_index[ind]])
elif self.split == 'Testing':
self.test_data = []
self.test_labels = []
for ind in range(len(test_index)):
self.test_data.append(self.data['data_pixel'][test_index[ind]])
self.test_labels.append(self.data['data_label'][test_index[ind]])
def __getitem__(self, index):
if self.split == 'Training':
img, target = self.train_data[index], self.train_labels[index]
elif self.split == 'Testing':
img, target = self.test_data[index], self.test_labels[index]
img = img[:,:,np.newaxis]
img = np.concatenate((img,img,img),axis=2)
img = Image.fromarray(img)
if self.transform is not None:
img = self.transform(img)
return img, target
def __len__(self):
if self.split == 'Training':
return len(self.train_data)
elif self.split == 'Testing':
return len(self.test_data)
if __name__ == '__main__':
import transforms as transforms
transform_train = transforms.Compose([
transforms.ToTensor(),
])
# transform_train = None
data = CK(split = 'Training', transform = transform_train)
for i in range(3):
print(data.__getitem__(i))
print(data.__len__())