微调Inception V3网络-对Satellite分类

  这篇博客主要是使用Keras框架微调Inception V3模型对卫星图片进行分类，并测试；

目录

• 1. 流程概述
• 2. 准备数据集
  • 2.1 Satellite数据集介绍
• 3. Inception V3网络
• 4. 训练
  • 4.1 基于Keras微调Inception V3网络
  • 4.2 Keras实时生成批量增强数据
  • 4.3 配置transfer learning & finetune
  • 4.4 执行训练
• 5. 测试
  • 5.1 对单张图片进行测试
• 6. 可视化分类界面
  • 6.1 交互界面设计
  • 6.2 后台核心代码：模型加载并分类
  • 6.3 交互界面效果

1. 流程概述

  微调Inception V3对卫星图片进行分类；整个流程可以大致分成四个步骤，如下：

• （1）Satellite数据集准备；
• （2）搭建Inception V3网络；
• （3）进行训练；
• （4）测试；

2. 准备数据集

2.1 Satellite数据集介绍

  用于实验训练与测试的数据集来自于《21个项目玩转深度学习：基于Tensorflow的实践详解》第三章中提供的实验卫星图片数据集；

  Satellite数据集目录结构如下：

# 其中共6类卫星图片，训练集总共4800张，每类800张；验证集共1200张，每类200张；
Satellite/
	train/
    	glacier/
        rock/
        urban/
        water/
        wetland/
        wood/
    validation/
    	glacier/
        rock/
        urban/
        water/
        wetland/
        wood/

3. Inception V3网络

  待补充；

4. 训练

4.1 基于Keras微调Inception V3网络

from keras.application.incepiton_v3 import InceptionV3, preprocess_input
from keras.layers import GlobalAveragePooling2D, Dense

#  基础Inception_V3模型，不包含全连接层
base_model = InceptionV3(weights='imagenet', include_top=False)
#  增加新的输出层
x = base_model.output
x = GlobalAveragePooling2D()(x) # 添加Global average pooling层
x = Dense(1024, activation='relu')(x)
predictions = Dense(6, activation='softmax')(x)

4.2 Keras实时生成批量增强数据

# keras实时生成批量增强数据
train_datagen = ImageDataGenerator(
    preprocessing_function=preprocess_input,  # 将每一张图片归一化到[-1,1]；数据增强后执行；
    rotation_range=30,
    width_shift_range=0.2,
    height_shift_range=0.2,
    shear_range=0.2,
    zoom_range=0.2,
    horizontal_flip=True,
)
val_datagen = ImageDataGenerator(
    preprocessing_function=preprocess_input,
    rotation_range=30,
    width_shift_range=0.2,
    height_shift_range=0.2,
    shear_range=0.2,
    zoom_range=0.2,
    horizontal_flip=True,
)

#  指定数据集路径并批量生成增强数据
train_generator = train_datagen.flow_from_directory(directory='satellite/data/train',
                                  target_size=(299, 299),#Inception V3规定大小
                                  batch_size=64)
val_generator = val_datagen.flow_from_directory(directory='satellite/data/validation',
                                target_size=(299,299),
                                batch_size=64)

4.3 配置transfer learning & finetune

from keras.optimizers import Adagrad

# transfer learning
def setup_to_transfer_learning(model,base_model):#base_model
    for layer in base_model.layers:
        layer.trainable = False
    model.compile(optimizer='adam', loss='categorical_crossentropy', metrics=['accuracy'])  # 配置模型，为下一步训练

# finetune
def setup_to_fine_tune(model,base_model):
    GAP_LAYER = 17  # max_pooling_2d_2
    for layer in base_model.layers[:GAP_LAYER+1]:
        layer.trainable = False
    for layer in base_model.layers[GAP_LAYER+1:]:
        layer.trainable = True
    model.compile(optimizer=Adagrad(lr=0.0001), loss='categorical_crossentropy', metrics=['accuracy'])

4.4 执行训练

# Step 1: transfer learning
setup_to_transfer_learning(model,base_model)
history_tl = model.fit_generator(generator=train_generator,
                    steps_per_epoch=75,  # 800
                    epochs=10,
                    validation_data=val_generator,
                    validation_steps=64,  # 12
                    class_weight='auto'
                    )
model.save('satellite/train_dir/satellite_iv3_tl.h5')

# Step 2: finetune
setup_to_fine_tune(model,base_model)
history_ft = model.fit_generator(generator=train_generator,
                                 steps_per_epoch=75,
                                 epochs=10,
                                 validation_data=val_generator,
                                 validation_steps=64,
                                 class_weight='auto')
model.save('satellite/train_dir/satellite_iv3_ft.h5')

5. 测试

5.1 对单张图片进行测试

# *-coding: utf-8 -*

"""
使用h5模型文件对satellite进行测试
"""
# ================================================================
import tensorflow as tf
import numpy as np
from skimage import io
from keras.models import load_model

def normalize(array):
    """对给定数组进行归一化

    Argument:
        array: array
            给定数组
    Return:
        array_norm: array
            归一化后的数组
    """
    array_flatten = array.flatten()
    array_mean = np.mean(array_flatten)
    mx = np.max(array_flatten)
    mn = np.min(array_flatten)
    array_norm = [(float(i) - array_mean) / (mx - mn) for i in array_flatten]

    return np.reshape(array_norm, array.shape)

def img_preprocess(image_path):
    """根据图片路径，对图片进行相应预处理

    Argument:
        image_path: str
            输入图片路径
    Return:
        image_data: array
            预处理好的图像数组
    """
    img_array = io.imread(image_path)
    img_norm = normalize(img_array)
    size = img_norm.shape
    image_data = np.reshape(img_norm, (1, size[0], size[1], 3))

    return image_data

def index_to_label(index):
    """将标签索引转换成可读的标签

    Argument:
        index: int
            标签索引位置
    Return:
        human_label: str
            人可读的标签
    """
    labels = ["glacier", "rock", "urban", "water", "wetland", "wood"]
    human_label = labels[index]

    return human_label

def classifier_satellite_byh5(image_path, model_file_path):
    """对给定单张图片使用训练好的模型进行分类

    Argument:
        image_path: str
            输入图片路径
        model_file_path: str
            训练好的h5模型文件名称
    Return:
        human_label: str
            人可读的图片标签
    """
    image_data = img_preprocess(image_path)
    # 加载模型文件
    model = load_model(model_file_path)
    predictions = model.predict(image_data)

    human_label = index_to_label(np.argmax(predictions))

    return human_label

def classifier_satellite_byh5_hci(image_path):
    """用于对从交互界面传来的图片进行分类

    Argument:
        image_path: str
    Return:
        human_label: str
            人可读的图片标签
    """
    # 模型文件，如果有新的模型需要修改
    model_file_path = "satellite/train_dir/models/satellite_iv3_ft.h5"

    image_data = img_preprocess(image_path)
    # 加载模型文件
    model = load_model(model_file_path)
    predictions = model.predict(image_data)

    human_label = index_to_label(np.argmax(predictions))

    return human_label

# 测试单张图片
if __name__ == "__main__":
    image_path = "satellite/data/train/glacier/40965_91335_18.jpg"
    model_file_path = "satellite/train_dir/models/satellite_iv3_ft.h5"

    human_label = classifier_satellite_byh5(image_path, model_file_path)
    print(human_label)

6. 可视化分类界面

6.1 交互界面设计

# encoding: utf-8
"""
交互界面：使用训练好的模型对卫星图片进行分类；
"""

from tkinter import *
import tkinter
import tkinter.filedialog
import os
import tkinter.messagebox
from PIL import Image, ImageTk
import test_satellite_bypb

# 窗口属性
root = tkinter.Tk()
root.title('Satellite图像分类')
root.geometry('800x600')

formatImg = ['jpg']

def resize(w, h, w_box, h_box, pil_image):
  # 对一个pil_image对象进行缩放，让它在一个矩形框内，还能保持比例

  f1 = 1.0*w_box/w # 1.0 forces float division in Python2
  f2 = 1.0*h_box/h
  factor = min([f1, f2])
  width = int(w*factor)
  height = int(h*factor)
  return pil_image.resize((width, height), Image.ANTIALIAS)

def showImg():
    img1 = entry_imgPath.get()  # 获取图片路径地址
    pil_image = Image.open(img1)    # 打开图片
    # 期望显示大小
    w_box = 400
    h_box = 400
    # 获取原始图像的大小
    w, h = pil_image.size
    pil_image_resized = resize(w, h, w_box, h_box, pil_image)

    # 把PIL图像对象转变为Tkinter的PhotoImage对象
    tk_image = ImageTk.PhotoImage(pil_image_resized)

    img = tkinter.Label(image=tk_image, width=w_box, height=h_box)
    img.image = tk_image
    img.place(x=50, y=150)

def choose_file():
    text_showClass.delete(0.0, END) # 清空输出结果文本框，在再次选择图片文件之前清空上次结果；
    selectFileName = tkinter.filedialog.askopenfilename(title='选择文件')  # 选择文件
    if selectFileName[-3:] not in formatImg:
        tkinter.messagebox.askokcancel(title='出错', message='未选择图片或图片格式不正确')   # 弹出错误窗口
        return
    else:
        e.set(selectFileName)  # 设置变量
        showImg()   # 显示图片

def ouputOfModel():
    # 完成识别，显示类别
    # 图片文件路径
    text_showClass.delete(0.0, END) # 清空上次结果文本框
    img_path = entry_imgPath.get()  # 获取所选择的图片路径地址

    # 判断是否存在改图片
    if not os.path.exists(img_path):
        tkinter.messagebox.askokcancel(title='出错', message='未选择图片文件或图片格式不正确')
    else:

        # 得到输出结果，以及相应概率
        human_label = test_satellite_bypb.classifier_satellite_img(img_path)
        # 通过训练的模型，计算得到相对应输出类别

        # 清空文本框中的内容，写入识别出来的类别
        text_showClass.config(state=NORMAL)
        text_showClass.insert('insert', '%s\n' % (human_label))

##################
# 窗口部件
##################

e = tkinter.StringVar() # 字符串变量

# label : 选择文件
label_selectImg = tkinter.Label(root, text='选择图片：')
label_selectImg.grid(row=0, column=0)

# Entry: 显示图片文件路径地址
entry_imgPath = tkinter.Entry(root, width=80, textvariable=e)
entry_imgPath.grid(row=0, column=1)

# Button: 选择图片文件
button_selectImg = tkinter.Button(root, text="选择", command=choose_file)
button_selectImg.grid(row=0, column=2)

# Button: 执行识别程序按钮
button_recogImg = tkinter.Button(root, text="开始识别", command=ouputOfModel)
button_recogImg.grid(row=0, column=3)

# Text: 显示结果类别文本框
text_showClass = tkinter.Text(root, width=20, height=1, font='18',)
text_showClass.grid(row=1, column=1)
text_showClass.config(state=DISABLED)

root.mainloop()

6.2 后台核心代码：模型加载并分类

# *-coding: utf-8 -*

"""
使用h5模型文件对satellite进行测试
"""
# ================================================================
import tensorflow as tf
import numpy as np
from skimage import io
from keras.models import load_model

def normalize(array):
    """对给定数组进行归一化

    Argument:
        array: array
            给定数组
    Return:
        array_norm: array
            归一化后的数组
    """
    array_flatten = array.flatten()
    array_mean = np.mean(array_flatten)
    mx = np.max(array_flatten)
    mn = np.min(array_flatten)
    array_norm = [(float(i) - array_mean) / (mx - mn) for i in array_flatten]

    return np.reshape(array_norm, array.shape)

def img_preprocess(image_path):
    """根据图片路径，对图片进行相应预处理

    Argument:
        image_path: str
            输入图片路径
    Return:
        image_data: array
            预处理好的图像数组
    """
    img_array = io.imread(image_path)
    img_norm = normalize(img_array)
    size = img_norm.shape
    image_data = np.reshape(img_norm, (1, size[0], size[1], 3))

    return image_data

def index_to_label(index):
    """将标签索引转换成可读的标签

    Argument:
        index: int
            标签索引位置
    Return:
        human_label: str
            人可读的标签
    """
    labels = ["glacier", "rock", "urban", "water", "wetland", "wood"]
    human_label = labels[index]

    return human_label

def classifier_satellite_byh5(image_path, model_file_path):
    """对给定单张图片使用训练好的模型进行分类

    Argument:
        image_path: str
            输入图片路径
        model_file_path: str
            训练好的h5模型文件名称
    Return:
        human_label: str
            人可读的图片标签
    """
    image_data = img_preprocess(image_path)
    # 加载模型文件
    model = load_model(model_file_path)
    predictions = model.predict(image_data)

    human_label = index_to_label(np.argmax(predictions))

    return human_label

def classifier_satellite_byh5_hci(image_path):
    """用于对从交互界面传来的图片进行分类

    Argument:
        image_path: str
    Return:
        human_label: str
            人可读的图片标签
    """
    # 模型文件，如果有新的模型需要修改
    model_file_path = "satellite/train_dir/models/satellite_iv3_ft.h5"

    image_data = img_preprocess(image_path)
    # 加载模型文件
    model = load_model(model_file_path)
    predictions = model.predict(image_data)

    human_label = index_to_label(np.argmax(predictions))

    return human_label

# 测试单张图片
if __name__ == "__main__":
    image_path = "satellite/data/train/glacier/40965_91335_18.jpg"
    model_file_path = "satellite/train_dir/models/satellite_iv3_ft.h5"

    human_label = classifier_satellite_byh5(image_path, model_file_path)
    print(human_label)

6.3 交互界面效果