使用TensorFlow的卷积神经网络识别手写数字（1）-预处理篇

　

功能：

　　将文件夹下的20*20像素黑白图片，根据重心位置绘制到28*28图片上，然后保存。经过预处理的图片有利于数字的准确识别。参见MNIST对图片的要求。

　　

　　此处可下载已处理好的图片：

　　https://files.cnblogs.com/files/hatemath/20-pixel-numbers.zip

　　https://files.cnblogs.com/files/hatemath/28-pixel-numbers.zip

 # encoding: utf-8
 import os

 from PIL import Image
 import numpy as np
 import cv2
 import matplotlib.pyplot as plt
 import matplotlib.cm as cm

 srcDir = '20-pixel-numbers'
 dstDir = '28-pixel-numbers'

 #显示图片
 def showImg(image):
     plt.imshow(image,cmap=cm.binary)
     plt.show()

 #按比例调整图片大小
 def resizeImage(image,width=None,height=None,inter=cv2.INTER_AREA):

     #获取图像尺寸
     (h,w) = image.shape[:2]
     if width is None and height is None:
         return image

     #高度算缩放比例

     if(w > h):
         newsize = (width,round(h / (w/width)))
     else:
         newsize = (round(w/ (h/height)), height)

         #print(newsize)

     # 缩放图像
     newimage = cv2.resize(image, newsize, interpolation=inter)
     return newimage

 #创建新的黑色图片
 def createBianryImage(bg=(0,0,0), width=28, height=28):

     channels = 1

     image = np.zeros((width,height,channels),np.uint8)#生成一个空灰度图像
     #cv2.rectangle(image,(0,0),(width,height),bg,1, -1)

     return image.reshape(width, height)

 #两个不同大小的图片合并
 def mergeImage(bg, fg, x, y):
     bgH, bgW = bg.shape[:2]
     fgH, fgW = fg.shape[:2]

     for i in range(fgH):
         for j in range(fgW):
             if(y+i < bgH and x+j < bgW):
                 #print('xx', y+i, x+j)
                 bg[y+i, x+j] = fg[i,j] # 这里可以处理每个像素点

     return bg

 # 求像素重心。传入二值图像，其中白色点算重量，黑色点为空
 def getBarycentre(image):

     h, w = image.shape[:2]

     sumWeightW = 0
     sumWeightH = 0

     count = 0

     for i in range(h):
         for j in range(w):
             if(image[i,j] > 128):
                 sumWeightW += j
                 sumWeightH += i
                 count += 1

     if(count == 0):
         count = 1

     print('getBarycentre: ', round(sumWeightW/count), round(sumWeightH/count) )
     return (round(sumWeightW/count), round(sumWeightH/count))

 def getFileList(strDir, strType='.png'):
     lstSrcFiles = []

     files = os.listdir(strDir)
     for file in files:
         if os.path.splitext(file)[1] == strType:
             lstSrcFiles.append(file)

     return lstSrcFiles

 # 读取指定目录下的图片文件，图片为黑白格式，长、宽的最大值为20像素。
 lstSrcFiles = getFileList(srcDir)
 print (lstSrcFiles)

 for file in lstSrcFiles:
     binary = cv2.imread(srcDir + '/' + file, cv2.IMREAD_GRAYSCALE)

     # 求像素重心
     bcW, bcH = getBarycentre(binary)

     # 叠加到28x28的黑色图片上
     xOffset = round(28/2 - bcW)
     yOffset = round(28/2 - bcH)    

     print('offset', xOffset, yOffset)

     # 另存为
     cv2.imwrite(dstDir + '/' + file,
                     mergeImage(createBianryImage(), binary, xOffset, yOffset))
                     #binary)