from keras.models import Sequential from keras.layers.normalization import BatchNormalization from keras.layers.convolutional import Conv2D from keras.layers.convolutional import MaxPooling2D from keras.layers.core import Activation from keras.layers.core import Flatten from keras.layers.core import Dropout from keras.layers.core import Dense from keras import backend as K from keras.preprocessing.image import ImageDataGenerator from keras.optimizers import Adam from keras.preprocessing import image from keras.preprocessing.image import img_to_array from sklearn.preprocessing import MultiLabelBinarizer from sklearn.model_selection import train_test_split from sklearn.preprocessing import OneHotEncoder from keras.utils import to_categorical from keras.applications import inception_v3 from keras.layers import GlobalAveragePooling2D from keras.models import Model import matplotlib.pyplot as plt import imutils import numpy as np import argparse import random import pickle import cv2 import os from PIL import Image import matplotlib matplotlib.use("Agg") # 获取该路径下所有图片 path = list(imutils.paths.list_images(r'C:\Users\Desktop\guangdong\train')) imagePaths = sorted(path) random.shuffle(imagePaths) name_dic = {
   '正常':'norm','不导电':'defect1','擦花':'defect2','横条压凹':'defect3','桔皮':'defect4','漏底':'defect5', '碰伤':'defect6','起坑':'defect7','凸粉':'defect8','涂层开裂':'defect9','脏点':'defect10','其他':'defect11'} # 将其他文件夹中，名称都改为其他 other_list_1 = os.listdir(r'C:\Users\Desktop\guangdong\train\guangdong_round1_train2_20180916\guangdong_round1_train2_20180916\瑕疵样本\其他') other_list = other_list_1[1:] other_dic = { '伤口':'其他', '划伤':'其他', '变形':'其他', '喷流':'其他', '喷涂碰伤':'其他', '打白点':'其他', '打磨印':'其他','拖烂':'其他', '杂色':'其他', '气泡':'其他', '油印':'其他', '油渣':'其他', '漆泡':'其他', '火山口':'其他', '碰凹':'其他', '粘接':'其他', '纹粗':'其他', '角位漏底':'其他', '返底':'其他', '铝屑':'其他', '驳口':'其他'} # 打印出name_dic里的英文部分，手动复制，再在每个后面添加‘：’及相应的数字 name_dic.values() digit_dir = {
   'norm':0, 'defect1':1, 'defect2':2, 'defect3':3, 'defect4':4, 'defect5':5, 'defect6':6, 'defect7':7, 'defect8':8, 'defect9':9, 'defect10':10, 'defect11':11} # 将图片resize成inception_v3需要的(299,299)大小，并转化成array labels = [] data =[] for imagePath in imagePaths: img = Image.open(imagePath) img = img.resize((299,299)) img = img_to_array(img) data.append(img) label_gbk = imagePath.split('\\')[-1].split('2')[0] if label_gbk in other_list: label_gbk = other_dic[label_gbk] label_english = name_dic[label_gbk] label = digit_dir[label_english] print(label_gbk,':',label_english,':',label) labels.append(label) # 像素归一化（有利于加速收敛） labels = np.array(labels) data = np.array(data, dtype="float") / 255.0 # 标签one-hot labels = to_categorical(labels) x_train, x_test, y_train, y_test = train_test_split(data, labels, test_size=0.2, random_state=42) # 数据增强 train_aug = ImageDataGenerator(rotation_range=25, width_shift_range=0.1,height_shift_range=0.1, shear_range=0.2, zoom_range=0.2, horizontal_flip=True, fill_mode="nearest",preprocessing_function=inception_v3.preprocess_input) # inception_v3基础模型，include_top=False就是要修改原模型的最后一层 base_model = inception_v3.InceptionV3(weights='imagenet',include_top=False) x = base_model.output x = GlobalAveragePooling2D()(x) x = Dense(units=1024,activation='relu')(x) predictions = Dense(units=12,activation='softmax')(x) model = Model(inputs=base_model.input, output=predictions) base_model.summary() model.summary() # 不训练基础层 for layer in base_model.layers: layer.trainable = False model.compile(optimizer='adam',loss='categorical_crossentropy',metrics=['accuracy']) # batch_size最好选2的n次方，参考的是内存格式 history_tl = model.fit_generator(generator=train_aug.flow(x=x_train,y=y_train,batch_size=32),validation_data=(x_test, y_test), steps_per_epoch=len(x_train)//32,epochs=10,verbose=1) model.save()