【人脸表情识别】pytorch处理CK+数据集:每日简讯

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：

(资料图)

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 imagesimport csvimport osimport numpy as npimport h5pyimport skimage.iock_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 informationdata_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_functionfrom PIL import Imageimport numpy as np import h5pyimport torch.utils.data as dataclass CK(data.Dataset):"""Args:train (bool, optional): If True, creates dataset from training set, otherwisecreates from test set.transform (callable, optional): A function/transforms that takes in an PIL imageand 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 = transformself.split = split # training set or test setself.fold = fold # the k-fold cross validationself.data = h5py.File("./H5File/CK+.h5","r",driver="core")number = len(self.data["data_label"]) # 981sum_number = [0, 135, 312, 387, 594, 678, 927, 981] # the sum of class numbertest_number = [12, 18, 9, 21, 9, 24, 6] # the number of each classtest_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 elementtest_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 arraysif 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, targetdef __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 transformstransform_train = transforms.Compose([transforms.ToTensor(),])# transform_train = Nonedata = CK(split = "Training", transform = transform_train)for i in range(3):print(data.__getitem__(i))print(data.__len__())