在 TensorFlow 或 python 中将多个图像转换为 CSV 文件

Question

我正在尝试获取多个图像的图像数据（即每个像素的灰度值），并将它们输出到一个 CSV 文件中，该文件包含每个图像的一行，每个像素的一列。最终我想对它们运行一个卷积神经网络，这样即使它们位于图像上的不同位置，我也可以对形状进行分类。

这是我生成形状的方法，如果你想自己做的话：

import numpy as np
import os
import tensorflow as tf
import tensorflow_datasets as tfds
from keras.models import Sequential
from keras.layers import Dense
from tensorflow.keras.layers import Dense, Conv2D, Dropout, Flatten, MaxPooling2D
from PIL import ImageDraw
from PIL import Image
import matplotlib.pyplot as plt
import numpy as np
from scipy import ndimage
from sklearn.preprocessing import LabelEncoder
import random
import imageio
import csv
import glob
print(tf.__version__)

# https://stackoverflow.com/questions/20747345/python-pil-draw-circle
#  https://diycomputerscienceandelectronics.wordpress.com/2017/04/28/pygame-and-python-continued-draw-shapes-randomly-at-random-positions-on-the-screen20170428orderasc/

training = np.array([])
n=1
for i in range(1000):

  img_size = 28
  rand_num1 = random.uniform(0,img_size-0.25*img_size)
  rand_num2 = random.uniform(0,img_size-0.3*img_size)
  rand_num3 = random.uniform(0,img_size-0.3*img_size)
  rand_num4 = random.uniform(0,img_size-0.3*img_size)
  rand_num5 = random.uniform(0,img_size-0.15*img_size)
  rand_num6 = random.uniform(0,img_size-0.15*img_size)
  
  image = Image.new('RGBA', (img_size,img_size))
  draw = ImageDraw.Draw(image)

  if n % 3 == 1:
    draw.ellipse((rand_num5, rand_num6, rand_num5+0.15*img_size, rand_num6+0.15*img_size), fill = 'black')
    training = np.append(training, [['circle']])
    image.save('circle%d.png' % i)
  if n % 3 == 2:
    draw.polygon([(rand_num3,rand_num2), (rand_num3+0.08*img_size, rand_num2+0.2*img_size), (rand_num3+0.25*img_size,rand_num2+0.05*img_size)], fill = "black")
    training = np.append(training, [['triangle']])
    image.save('triangle%d.png' % i)
  if n % 3 == 0:
    draw.rectangle(((rand_num1, rand_num4), (rand_num1+0.15*img_size, rand_num4+0.25*img_size)), fill="black")
    training = np.append(training, [['rectangle']])
    image.save('rectangle%d.png' % i)
  n = n + 1

testing = np.array([])

for i in range(1000):

  img_size = 28
  rand_num1 = random.uniform(0,img_size-0.25*img_size)
  rand_num2 = random.uniform(0,img_size-0.3*img_size)
  rand_num3 = random.uniform(0,img_size-0.3*img_size)
  rand_num4 = random.uniform(0,img_size-0.3*img_size)
  rand_num5 = random.uniform(0,img_size-0.15*img_size)
  rand_num6 = random.uniform(0,img_size-0.15*img_size)
  
  image = Image.new('RGBA', (img_size,img_size))
  draw = ImageDraw.Draw(image)

  if n % 3 == 1:
    draw.ellipse((rand_num5, rand_num6, rand_num5+0.15*img_size, rand_num6+0.15*img_size), fill = 'blue', outline ='blue')
    testing = np.append(testing, [['circle']])
    
  if n % 3 == 2:
    draw.polygon([(rand_num3,rand_num2), (rand_num3+0.08*img_size, rand_num2+0.2*img_size), (rand_num3+0.25*img_size,rand_num2+0.05*img_size)], fill = (255,0,0))
    testing = np.append(testing, [['triangle']])
    
  if n % 3 == 0:
    draw.rectangle(((rand_num1, rand_num4), (rand_num1+0.15*img_size, rand_num4+0.25*img_size)), fill="black")
    testing = np.append(testing, [['rectangle']])
    
  n = n + 1

我知道这些链接接近我想要的，但不完全是： How can I convert a png to a dataframe for python? Converting images to csv file in python

当我尝试使用这些解决方案时，我没有得到任何像素数据——只有零——而且我只能让程序对一张图像起作用，而不是我数据库中的 1000 张。

这是我尝试过的：

import numpy as np
import cv2 
import csv 

for filen in glob.glob('*.png'):
#img_file = Image.open(filen)
  img = cv2.imread(filen, 0) # load grayscale image. Shape (28,28)

  flattened = img.flatten() # flatten the image, new shape (784,)

  flattened = np.insert(flattened, 0, 0) # insert the label at the beginning of the array, in this case we add a 0 at the index 0. Shape (785,0)


  #create column names 
  column_names = []
  column_names.append("label")
  [column_names.append("pixel"+str(x)) for x in range(0, 784)] # shape (785,0)

  # write to csv 
  with open('custom_test.csv', 'w') as file:
      writer = csv.writer(file, delimiter=';')
      writer.writerows([column_names]) # dump names into csv
      writer.writerows([flattened]) # add image row 
      # optional: add addtional image rows

和

def createFileList(myDir, format='.pgn'):
  fileList = []
  print(myDir)
  for root, dirs, files in os.walk(myDir, topdown=False):
      for name in files:
          if name.endswith(format):
              fullName = os.path.join(root, name)
              fileList.append(fullName)
  return fileList

for filen in glob.glob('*.png'):
    print(filen) 
    img_file = Image.open(filen)
    # img_file.show()

    # get original image parameters...
    width, height = img_file.size
    format = img_file.format
    mode = img_file.mode

    # Make image Greyscale
    #img_grey = img_file.convert('L')
    #img_grey.save('result.png')
    #img_grey.show()

    
    # Save Greyscale values
    value = np.asarray(img_file.getdata(), dtype=np.int).reshape((28, 28))
    value = value.flatten()
    train_example.append(value)
    print(value)
    with open("training_data.csv", 'w') as f:
        writer = csv.writer(f)
        writer.writerow(value) 
print(train_example)

Answer 1

更新：我认为出于某种原因，我的像素信息被保存到 Alpha 通道中。我只是从那里得到你的灰度图像，所以我写了

img_grey = img_file.getchannel("A")

然后从 img_grey 获取我的 np 数组。

另外，为了确保每个新的输入图像都不会在 CSV 文件中被覆盖（让我在 CSV 文件中有一行只有输入中的最后一个图像），我更改了 @ 中的 'w' 987654323@ 到'a'。