人脸剪裁，将图片转csv,表情识别的步骤

参考：数据科学竞赛-人脸表情识别

将文件夹里的图片===》符合要求的图片（只包含人脸，且转化为一定的大小）

from config import data_folder, generate_folder
import os
import numpy as np
import cv2
from PIL import Image
from glob import glob
from tqdm import tqdm
import pandas as pd

import face_recognition


if not os.path.exists(generate_folder):
    os.mkdir(generate_folder)


def scale_img(img, scale_size=139):
    h, w = img.shape[:2]
    if h > w:
        new_h, new_w = scale_size * h / w, scale_size
    else:
        new_h, new_w = scale_size, scale_size * w / h

    new_h, new_w = int(new_h), int(new_w)
    img = cv2.resize(img, (new_w, new_h))
    top = None
    left = None
    if h == w:
        return img
    elif h < w:
        if new_w > scale_size:
            left = np.random.randint(0, new_w - scale_size)
        else:
            left = 0
        top = 0

    elif h > w:
        if new_h > scale_size:
            top = np.random.randint(0, new_h - scale_size)
        else:
            top = 0
        left = 0
    img = img[top: top + scale_size, left: left + scale_size]
    return img


def preprocess():
    categories = os.listdir(data_folder)
    for category in categories:
        in_path = os.path.join(data_folder, category)
        out_path = os.path.join(generate_folder, category + '_face')
        if not os.path.exists(out_path):
            os.mkdir(out_path)
        for file in glob(in_path + '/*.jpg'):
            file_name = file.split('\\')[-1]  # Linux系统需要修改这一项file_name = file.split('/')[-1]
            print(file_name)
            img = face_recognition.load_image_file(file)
            if max(img.shape) > 2000:
                if img.shape[0] > img.shape[1]:
                    img = cv2.resize(img, (2000, int(2000 * img.shape[1] / img.shape[0])))
                else:
                    img = cv2.resize(img, (int(2000 * img.shape[0] / img.shape[1]), 2000))
            locations = face_recognition.face_locations(img)  # 人脸检测，大部分为单个，但也有多个检测结果
            if len(locations) <= 0:
                print("no face")
            else:
                for i, (a, b, c, d) in enumerate(locations):
                    image_split = img[a:c, d:b, :]
                    image_split = scale_img(image_split)
                    Image.fromarray(image_split).save(os.path.join(out_path, file_name + '_{}.png'.format(i)))

||
||

将这里的图片都转化为csv文件，两列，一列文件名，一列标签

def generate_desc_csv(folder):
    categories = os.listdir(folder)
    file_id = []
    label = []
    for category in tqdm(categories):
        # print(type(category))
        images = glob(os.path.join(folder, category) + '/*.png')  # 这里可以观察一下数据集，发现图片均为jpg格式，正则匹配比较简单
        for img in images:
            file_id.append(img)
            label.append(category)
    df_description = pd.DataFrame({'file_id': file_id, 'label': label})
    df_description.to_csv('../data/description.csv', encoding='utf8', index=False)  # 落地这个csv文件是为了更符合常见的数据集说明文件

查看csv文件，可以看到相应标签的数量,可以观察数据

df_data=pd.read_csv('../data/description.csv',encoding='utf-8')
print(type(df_data['label']))
print(df_data.sample(5))

print(df_data['label'].value_counts())
df_data['label'].value_counts().plot(kind='bar')
# plt.show()

随机采样10张图片展示并显示标签

datasize=len(df_data)
print('datasize',datasize)
sample_index=random.sample(list(range(datasize)),10)
print(sample_index)

plt.figure(figsize=(18,6))
for i in range(10):
    plt.subplot(2,5,i+1)
    img=cv2.imread(df_data['file_id'][sample_index[i]])
    img=cv2.cvtColor(img,cv2.COLOR_BGR2RGB)
    plt.imshow(img)
    label=str(df_data['label'][sample_index[i]])
    plt.title('label {}'.format(label))

plt.show()

数据增广

"""
Author: Zhou Chen
Date: 2020/1/8
Desc: desc
"""
from keras.preprocessing.image import ImageDataGenerator
import pandas as pd


class DataSet(object):

    def __init__(self, root_folder):
        self.folder = root_folder
        self.df_desc = pd.read_csv(self.folder + 'description.csv', encoding="utf8")
        print('the total length of dataset: ',len(self.df_desc))

    def get_generator(self, batch_size=32, da=True):
        if da:
            # 数据增强,多少数据用于训练集集，水平反转()
            train_gen = ImageDataGenerator(rescale=1 / 255., validation_split=0.25, horizontal_flip=True)
        else:
            train_gen = ImageDataGenerator(rescale=1 / 255., validation_split=0.25, horizontal_flip=False)
        img_size = (64, 64)
        train_generator = train_gen.flow_from_dataframe(dataframe=self.df_desc, #一列为图像的文件名，另一列为图像的类别
                                                        directory='.', #目标目录的路径
                                                        x_col='file_id',
                                                        y_col='label',
                                                        batch_size=batch_size,
                                                        class_mode='categorical',#将是 2D one-hot 编码标签
                                                        target_size=img_size,
                                                        color_mode='grayscale', #图像是否被转换成 1 或 3 个颜色通道。
                                                        subset='training')
        valid_generator = train_gen.flow_from_dataframe(dataframe=self.df_desc,
                                                        directory=".",
                                                        x_col="file_id",
                                                        y_col="label",
                                                        batch_size=batch_size,
                                                        class_mode="categorical",
                                                        target_size=img_size,
                                                        color_mode='grayscale',
                                                        subset='validation')
        return train_generator, valid_generator

if __name__=='__main__':
    ds = DataSet('../data/')
    train_generator, valid_generator = ds.get_generator(batch_size=64)
    print('the length of train dataset',train_generator.n)
    print('the size of one batch : ',train_generator.batch_size)

result:

train_gen.flow_from_dataframe出错?
重新安装keras.processing
pip install +具体位置/keras.processing.zip

keras.processing.zip
链接：https://pan.baidu.com/s/1Vo3lPrt4itjrzqudgugfwg
提取码：qigx

模型构建

from tensorflow.keras.layers import Input, Conv2D, PReLU, MaxPooling2D, Dense, Flatten, Dropout, GlobalAveragePooling2D
from tensorflow.keras.models import Model
from tensorflow.keras.applications import ResNet50


def CNN(input_shape=(224, 224, 3), n_classes=5):
    # input
    input_layer = Input(shape=input_shape)
    x = Conv2D(32, (1, 1), strides=1, padding='same', activation='relu')(input_layer)
    # block1
    x = Conv2D(64, (3, 3), strides=1, padding='same')(x)
    x = PReLU()(x)
    x = Conv2D(64, (5, 5), strides=1, padding='same')(x)
    x = PReLU()(x)
    x = MaxPooling2D(pool_size=(2, 2), strides=2)(x)
    # fc
    x = Flatten()(x)
    x = Dense(2048, activation='relu')(x)
    x = Dropout(0.5)(x)
    x = Dense(1024, activation='relu')(x)
    x = Dropout(0.5)(x)
    x = Dense(n_classes, activation='softmax')(x)

    model = Model(inputs=input_layer, outputs=x)
    return model


def ResNet_pretrained(input_shape=(224, 224, 3), n_classes=5):
    input_layer = Input(shape=input_shape)
    densenet121 = ResNet50(include_top=False, weights=None, input_tensor=input_layer)
    x = GlobalAveragePooling2D()(densenet121.output)
    x = Dropout(0.5)(x)
    x = Dense(n_classes, activation='softmax')(x)

    model = Model(input_layer, x)
    return model

if __name__=='__main__':
    resnet_model=ResNet_pretrained()
    resnet_model.summary()

训练模型

from data import DataSet
from model import CNN, ResNet_pretrained
from tensorflow.keras.callbacks import ModelCheckpoint, EarlyStopping
from tensorflow.keras.optimizers import Adam
from utils import plot_history
import os

if not os.path.exists("../models/"):
    os.mkdir("../models/")


# 数据
ds = DataSet('../data/')
train_generator, valid_generator = ds.get_generator(batch_size=64)
img_shape = (64, 64, 1)
# 模型
model_cnn = CNN(input_shape=img_shape)
model_resnet = ResNet_pretrained(input_shape=img_shape)
# 训练
optimizer_cnn = Adam(lr=3e-4)
optimizer_resnet = Adam(lr=3e-4)
callbacks_cnn = [
    ModelCheckpoint('../models/cnn_best_weights.h5', monitor='val_loss', save_best_only=True, verbose=1, save_weights_only=True),
    # EarlyStopping(monitor='val_loss', patience=5)
]

callbacks_resnet = [
    ModelCheckpoint('../models/resnet_best_weights.h5', monitor='val_loss', save_best_only=True, verbose=1, save_weights_only=True),
    # EarlyStopping(monitor='val_loss', patience=5)
]

model_cnn.compile(optimizer=optimizer_cnn, loss='categorical_crossentropy', metrics=['accuracy'])
model_resnet.compile(optimizer=optimizer_resnet, loss='categorical_crossentropy', metrics=['accuracy'])

epochs = 20
history_cnn = model_cnn.fit_generator(train_generator,
                              steps_per_epoch=train_generator.n//train_generator.batch_size,
                              validation_data=valid_generator,
                              validation_steps=valid_generator.n//valid_generator.batch_size,
                              epochs=epochs,
                              callbacks=callbacks_cnn
                             )
history_resnet = model_cnn.fit_generator(train_generator,
                              steps_per_epoch=train_generator.n//train_generator.batch_size,
                              validation_data=valid_generator,
                              validation_steps=valid_generator.n//valid_generator.batch_size,
                              epochs=epochs,
                              callbacks=callbacks_resnet
                             )

plot_history([history_cnn, history_resnet])

画出各个模型分别在训练集，验证集上的精度

import matplotlib.pyplot as plt
import numpy as np
plt.style.use('fivethirtyeight')


def plot_history(his):
    cnn_his = his[0].history
    resnet_his = his[1].history
    plt.figure(figsize=(18, 6))
    plt.subplot(1, 2, 1)
    plt.plot(np.arange(len(cnn_his['accuracy'])), cnn_his['accuracy'], label="training accuracy")
    plt.plot(np.arange(len(cnn_his['val_accuracy'])), cnn_his['val_accuracy'], label="validation accuracy")
    plt.title("CNN")
    plt.legend(loc=0)

    plt.subplot(1, 2, 2)
    plt.plot(np.arange(len(resnet_his['accuracy'])), resnet_his['accuracy'], label="training accuracy")
    plt.plot(np.arange(len(resnet_his['val_accuracy'])), resnet_his['val_accuracy'], label="validation accuracy")
    plt.title("ResNet50")
    plt.legend(loc=0)

    plt.savefig("../assets/his.png")
    plt.show()