目录

• 数据处理与数据集制备
• • 一、图片批量重命名
  • 二、数据增强+高斯加噪
  • 三、数据集制备

一、图片批量重命名

文件夹结构：

• source
  • animals
    • cat
    • cattle
  • • dog
    • horse
    • pig
  • fruits
    • apple
    • banana
    • durian
    • grape
    • orange
  • vehicles
    • bus
    • car
    • plane
    • ship
    • train

• 实现代码：

  import os
  import cv2 as cv
  
  sourcePath  = '../../../DataSet/source/'
  
  animalPath  = sourcePath + 'animals/'
  fruitPath   = sourcePath + 'fruits/'
  vehiclePath = sourcePath + 'vehicles/'
  
  
  # 水果图片重命名
  for folder_list in os.listdir(fruitPath):
  
      count = 1
  
      # 读取每个目录下的子文件夹
      subfolder = fruitPath + folder_list + '/'
  
      for file in os.listdir(subfolder):
  
          old_name = subfolder + file
          new_name = subfolder + "%s_%d.jpg" % (folder_list, count)
          print(old_name, "====>", new_name)
  
          os.rename(old_name, new_name)
          count = count + 1
  
  
  # 动物图片重命名
  for folder_list in os.listdir(animalPath):
  
      count = 1
  
      # 读取每个目录下的子文件夹
      subfolder = animalPath + folder_list + '/'
  
      for file in os.listdir(subfolder):
  
          old_name = subfolder + file
          new_name = subfolder + "%s_%d.jpg" % (folder_list, count)
  
          print(old_name, "====>", new_name)
  
          os.rename(old_name, new_name)
          count = count + 1
  
  
  # 交通图片重命名
  for folder_list in os.listdir(vehiclePath):
  
      count = 1
  
      # 读取每个目录下的子文件夹
      subfolder = vehiclePath + folder_list + '/'
  
      for file in os.listdir(subfolder):
  
          old_name = subfolder + file
          new_name = subfolder + "%s_%d.jpg" % (folder_list, count)
  
          print(old_name, "====>", new_name)
  
          os.rename(old_name, new_name)
          count = count + 1
  
  
  
  

实现效果：

二、数据增强+高斯加噪

比赛的时候采用摄像头拍摄图像识别，拍摄的图像包含很多噪点，而原数据集为无噪声的原图。例如一张狗的照片拍摄时受到室内光线影响，有很多红色噪点，就容易被识别为苹果。去年出现过这种情况，因此今年针对解决，对数据集进行加噪处理。

• 实现代码：

  # 图像加噪
  # 实际拍摄的图片会存在很多噪声
  # 与计算机原图差距很大
  # 因此需要对原图进行加噪处理
  
  import numpy as np
  import cv2   as cv
  import os
  from keras.preprocessing.image import ImageDataGenerator
  from keras.preprocessing.image import img_to_array
  from keras.preprocessing.image import load_img
  
  
  # 加噪函数
  def Gasuss_Noise(image, mean=0, var=0.001):
      '''
          添加高斯噪声
          image: 原始图像
          mean : 均值
          var :  方差, 越大，噪声越大
      '''
      image = cv.imread(image)
      image = np.array(image/255, dtype=float)                                                                            # 将原始图像的像素值进行归一化，除以255使得像素值在0-1之间
      noise = np.random.normal(mean, var ** 0.5, image.shape)                                                             # 创建一个均值为mean，方差为var呈高斯分布的图像矩阵
      out = image + noise                                                                                                 # 将噪声和原始图像进行相加得到加噪后的图像
      if out.min() < 0:
          low_clip = -1.
      else:
          low_clip = 0.
      out = np.clip(out, low_clip, 1.0)                                                                                   # clip函数将元素的大小限制在了low_clip和1之间了，小于的用low_clip代替，大于1的用1代替
      out = np.uint8(out * 255)                                                                                             # 解除归一化，乘以255将加噪后的图像的像素值恢复
      # cv.imshow("gasuss", out)
      noise = noise * 255
      return [noise, out]
  
  
  # 图像增强函数
  def Date_Enhancement(img_input_path, img_output_path):
      image = load_img(img_input_path)
  
      im1 = image.point(lambda p: p * 0.6)
      # im1.show()
  
      im1 = img_to_array(im1)  # 图像转为数组
      im1 = np.expand_dims(im1, axis=0)  # 增加一个维度
  
      img_dag = ImageDataGenerator(
          rotation_range=10,
          width_shift_range=0.001,   # horizen offset
          height_shift_range=0.001,  # vertical offset
          shear_range=0.02,
          # zoom_range=[0.6, 0.9],
          brightness_range = [0.9, 1.1],
          horizontal_flip=False,    # 水平翻转
          fill_mode="constant", cval=40
      )  # 旋转，宽度移动范围，高度移动范围，裁剪范围，水平翻转开启，填充模式
  
      img_generator = img_dag.flow(im1,
                                   batch_size=1,
                                   save_to_dir= img_output_path,
                                   save_prefix= "image",
                                   save_format= "jpg")  # 测试一张图像bath_size=1
      count = 0  # 计数器
  
      for raw_pic_path in img_generator:
          count += 1
          if count == 1:  # 生成多少个样本后退出
              break
  
  
  ####################################################################
  
  # 原地址
  sourcePath  = '../../../DataSet/source/'
  animalPath  = sourcePath + 'animals/'
  fruitPath   = sourcePath + 'fruits/'
  vehiclePath = sourcePath + 'vehicles/'
  
  # 中间地址
  # 用于保存第一步处理后的图片
  middlePath        = '../../../DataSet/middle/'
  animalMiddlePath  = middlePath + 'animals/'
  fruitMiddlePath   = middlePath + 'fruits/'
  vehicleMiddlePath = middlePath + 'vehicles/'
  
  # 目标地址
  # 保存第二步处理后的图片数据
  processPath        = '../../../DataSet/process/'
  animalProcessPath  = processPath + 'animals/'
  fruitProcessPath   = processPath + 'fruits/'
  vehicleProcessPath = processPath + 'vehicles/'
  
  
  for folder_name in os.listdir(fruitPath):
  
      rename_count = 1
  
      # 读取每个目录下的子文件夹
      sub_folder = fruitPath + folder_name + '/'
      middle_save_path = fruitMiddlePath + folder_name + '/'
  
      # 数据增强
      for raw_pic in os.listdir(sub_folder):
          raw_pic_path = sub_folder + raw_pic
          for i in range(0, 10):
              Date_Enhancement(raw_pic_path, middle_save_path)
  
      print(folder_name, "Enhance Done")
  
      # 高斯加噪
      for middle_pic in os.listdir(middle_save_path):
          middle_save_name = middle_save_path + middle_pic
          final_save_path = fruitProcessPath + folder_name + '/'
          final_save_name = final_save_path + "%s_%d.jpg" % (folder_name, rename_count)
  
          noise, out = Gasuss_Noise(middle_save_name, mean=0, var=0.002)
  
          cv.imwrite(final_save_name, out)
          rename_count = rename_count + 1
  
      print(folder_name, "Process Done")
  
  
  for folder_name in os.listdir(animalPath):
  
      rename_count = 1
  
      # 读取每个目录下的子文件夹
      sub_folder = animalPath + folder_name + '/'
      middle_save_path = animalMiddlePath + folder_name + '/'
  
      # 数据增强
      for raw_pic in os.listdir(sub_folder):
          raw_pic_path = sub_folder + raw_pic
          for i in range(0, 10):
              Date_Enhancement(raw_pic_path, middle_save_path)
  
      print(folder_name, "Enhance Done")
  
      # 高斯加噪
      for middle_pic in os.listdir(middle_save_path):
          middle_save_name = middle_save_path + middle_pic
          final_save_path = animalProcessPath + folder_name + '/'
          final_save_name = final_save_path + "%s_%d.jpg" % (folder_name, rename_count)
  
          noise, out = Gasuss_Noise(middle_save_name, mean=0, var=0.002)
  
          cv.imwrite(final_save_name, out)
          rename_count = rename_count + 1
  
      print(folder_name, "Process Done")
  
  
  for folder_name in os.listdir(vehiclePath):
  
      rename_count = 1
  
      # 读取每个目录下的子文件夹
      sub_folder = vehiclePath + folder_name + '/'
      middle_save_path = vehicleMiddlePath + folder_name + '/'
  
      # 数据增强
      for raw_pic in os.listdir(sub_folder):
          rename_count = 1
          raw_pic_path = sub_folder + raw_pic
          for i in range(0, 10):
              Date_Enhancement(raw_pic_path, middle_save_path)
  
      print(folder_name, "Enhance Done")
  
      # 高斯加噪
      for middle_pic in os.listdir(middle_save_path):
          middle_save_name = middle_save_path + middle_pic
          final_save_path = vehicleProcessPath + folder_name + '/'
          final_save_name = final_save_path + "%s_%d.jpg" % (folder_name, rename_count)
  
          noise, out = Gasuss_Noise(middle_save_name, mean=0, var=0.002)
  
          cv.imwrite(final_save_name, out)
          rename_count = rename_count + 1
  
      print(folder_name, "Process Done")
  

三、数据集制备

【说明】python访问文件/文件夹的顺序是按照首字母的排序进行的，于是标签的顺序为：

• 实现代码：

  import os
  import cv2 as cv
  import numpy as np
  
  make = True
  check = True
  
  # 图片地址
  picPath        = '../../../DataSet/process/'
  
  if __name__ == "__main__":
  
      if make:
          all_data = []
          all_label = []
  
          i = -1
          for fruit_animal_vehicle in os.listdir(picPath):
              for apple_cat_bus in os.listdir(picPath + fruit_animal_vehicle + '/'):
                  i = i + 1
                  for pic in os.listdir(picPath + fruit_animal_vehicle + '/' + apple_cat_bus + '/'):
  
                      extension = os.path.splitext(pic)[-1]
                      if extension == '.jpg':
                          img = cv.imread(picPath + fruit_animal_vehicle + '/' + apple_cat_bus + '/' + pic)  # 读取图片数据到img   ##连接两个路径名组件  #  cv2.imread 读取的图片数据是 BGR 格式；
                          try:                                                   # 如果try发生异常
                              img = cv.resize(img, (32, 32))[..., (2, 1, 0)]        # BGR  2  RGB
                              all_data.append(img)                              # .append方法在数组尾部加上img
                              all_label.append(i)                               # .append方法
                          except:                                                # 执行except
                              continue
  
          all_data  = np.asarray(all_data)
          all_label = np.asarray(all_label)
  
          np.save("../../../DataSet/dataSet/pic", all_data)
          np.save("../../../DataSet/dataSet/label", all_label)
  
      if check:
          x = np.load("../../../DataSet/dataSet/pic.npy")
          y = np.load("../../../DataSet/dataSet/label.npy")
  
          label = ["cat", "cattle", "dog", "horse", "pig"] \
                + ["apple", "banana", "durian", "grape", "orange"] \
                + ["bus", "car", "plane", "ship", "train"]
  
          count = 0
          for d, idx in zip(x, y):                                             ## 将对象中对应的元素打包成一个个元组，然后返回由这些元组组成的列表
              print("Class %s %d" % (label[idx], count))
              d = cv.resize(d, (395, 395))[..., (2, 1, 0)]
              cv.imshow("img", d)
              count = count + 1
              cv.waitKey(0)