使用TensorFlow的卷積神經網絡識別手寫數字（1）-預處理篇

　

功能：

　　將文件夾下的20*20像素黑白圖片，根據重心位置繪制到28*28圖片上，然后保存。經過預處理的圖片有利於數字的准確識別。參見MNIST對圖片的要求。

 　　

　　此處可下載已處理好的圖片：

　　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'

#顯示圖片
def showImg(image):
    plt.imshow(image,cmap=cm.binary)
    plt.show()

#按比例調整圖片大小
def resizeImage(image,width=None,height=None,inter=cv2.INTER_AREA):
    
    #獲取圖像尺寸
    (h,w) = image.shape[:2]
    if width is None and height is None:
        return image
    
    #高度算縮放比例

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

        #print(newsize)
    

        
    # 縮放圖像
    newimage = cv2.resize(image, newsize, interpolation=inter)
    return newimage

#創建新的黑色圖片 
def createBianryImage(bg=(0,0,0), width=28, height=28):

    channels = 1

    image = np.zeros((width,height,channels),np.uint8)#生成一個空灰度圖像
    #cv2.rectangle(image,(0,0),(width,height),bg,1, -1)
    
    return image.reshape(width, height)

#兩個不同大小的圖片合並
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] # 這里可以處理每個像素點
        
    return bg

# 求像素重心。傳入二值圖像，其中白色點算重量，黑色點為空
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
            
    
# 讀取指定目錄下的圖片文件，圖片為黑白格式，長、寬的最大值為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)

    # 疊加到28x28的黑色圖片上
    xOffset = round(28/2 - bcW)
    yOffset = round(28/2 - bcH)    

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