基于Python和TensorFlow平台的神经网络进行图像识别

本文主要是记录课程设计，使用Python和TensorFlow平台进行图像训练识别。

识别结果：

代码：

#coding=utf-8
import os
import sys
from PIL import Image #使用第三方包进行图像处理
import numpy as np
import tensorflow.contrib.keras as k
import scipy.spatial.distance as distance
import pandas as pd

def resizeImage(inputImage, maxWidth, maxHeight):
    originalWidth, originalHeight = inputImage.size

    f1 = 1.0*maxWidth / originalWidth
    f2 = 1.0*maxHeight / originalHeight
    factor = min([f1,f2])
    width = int(originalWidth*factor)
    height = int(originalHeight*factor)
    return  inputImage.resize((width,height),Image.ANTIALIAS)

ifRestartT = False
roundCount = 20
learnRate = 0.01
trainDir="../imageData/animals"  #用于指定训练数据所在的目录
trainResultPath = "./imageclassifySave/"     #用于指定训练数据保存目录
optimizerT = "RMSProp"
lossT = "categorical_crossentropy"
predictFile = None

#读取meta.txt文件内容
metaData=pd.read_csv(trainDir+"meta.txt",header=None).as_matrix()

maxTypes = len(metaData)
print("maxTypes: %d" % maxTypes)

argt = sys.argv[1:]
print("argt:%s" % argt)

for v in argt:
    if v == "-restart":
        ifRestartT = True
    if v.startswith("-round="):
        roundCount = int(v[len("-round="):])
    if v.startswith("-learnrate="):
        learnRate = float(v[len("-learnrate="):])
    if v.startswith("-dir="):
        trainDir = v[len("-dir="):]
    if v.startswith("-optimizer="):
        optimizerT = v[len("-optimizer="):]
    if v.startswith("-loss="):
        lossT = v[len("-loss="):]
    if v.startswith("-predict="):
        #predictFile="imageData/test/"+v[len("-predict="):]
        predictFile=v[len("-predict="):]
print("predict file: %s" % predictFile)
xData =[]
yTrainData = []
fnData = []
predictAry = []

listt = os.listdir(trainDir)

lent = len(listt)

for i in range(lent):
    v = listt[i]
    if v.endswith(".jpg"):
        print("processing %s ..." % v)
        img = Image.open(trainDir + "/"+v)
        w,h =img.size

        img1 = resizeImage(img,512,512)
        img2 = Image.new("RGB",(512,512),color = "white")

        w1,h1 = img1.size
        img2 = Image.new("RGB",(512,512),color ="white")
        img2.paste(img1,box=(int((512 - w1)/2),int((512 -h1)/2)))
        xData.append(np.matrix(list(img2.getdata())))
        tmpv = np.full([maxTypes],fill_value=0)
        tmpv[int(v.split(sep="_")[0]) - 1] = 1
        yTrainData.append(tmpv)
        fnData.append(trainDir+v)

rowCount = len(xData)
print("rowCount:%d" % rowCount)
#转换XData,yTrainData,fnCount为合适的形态
xData = np.array(xData)
xData = np.reshape(xData,(-1,512,512,3))
yTrainData = np.array(yTrainData)

fnData = np.array(fnData)

if(ifRestartT is False) and os.path.exists(trainResultPath+".h5"):
    #载入保存的模型和可变参数
    print("Loading...")
    model = k.models.load_model(trainResultPath+".h5")
    model.load_weights(trainResultPath+"wb.h5")
else:
    model=k.models.Sequential()
    

    model.add(k.layers.Conv2D(filters=4,kernel_size=(3,3),input_shape=(512,512,3),data_format="channels_last",activation="relu"))
    model.add(k.layers.Conv2D(filters=3,kernel_size=(3,3),data_format="channels_last",activation="relu"))
    model.add(k.layers.Conv2D(filters=2,kernel_size=(2,2),data_format="channels_last",activation="selu"))

    model.add(k.layers.Flatten())
    model.add(k.layers.Dense(256,activation='tanh'))
    model.add(k.layers.Dense(64,activation='sigmoid'))

    #按分类数进行softmax分类
    model.add(k.layers.Dense(maxTypes,activation='softmax'))
    model.compile(loss=lossT,optimizer=optimizerT,metrics=['accuracy'])

if predictFile is not None:
    #先对已有训练数据执行一遍预测，以便后面做图片相似度比对
    print("preparing...")
    predictAry = model.predict(xData)
    print("processing %s ..." % predictFile)
    img = Image.open(predictFile)

    #下面对新输入图片进行预测
    img1=resizeImage(img,512,512)
    w1, h1=img1.size
    img2=Image.new("RGB",(512,512),color="white")
    img2.paste(img1,box=(int((512 - w1)/2),int((512 - h1)/2)))

    xi = np.matrix(list(img2.getdata()))
    xi1 = np.array(xi)
    xin = np.reshape(xi1,(-1,512,512,3))
    resultAry = model.predict(xin)
    print("x:%s, y:%s"%(xin,resultAry))
    #找出预测结果中最大可能的概率及其对应的编号
    maxIdx = -1
    maxPercent = 0

    for i in range(maxTypes):
        if resultAry[0][i]>maxPercent:
            maxPercent = resultAry[0][i]
            maxIdx = i
    #将新图片的预测结果与训练图片的预测结果逐一比对，找出相似度最高的
    minDistance = 200
    minIdx = -1
    minFile = ""

    for i in range(rowCount):
        dist = distance.cosine(resultAry[0],predictAry[i])
        if dist<minDistance:
            minDistance = dist
            minIdx = i
            minFile = fnData[i]
    print("图片类型: %s,推断正确概率: %10.6f%%, 最相识文件:%s, 相似度: %10.6f%%" %(metaData[maxIdx][1],maxPercent*100,minFile.split("\\")[-1],(1-minDistance)*100))
    sys.exit(0)
model.fit(xData,yTrainData,epochs=roundCount,batch_size=lent,verbose=2)
print("saving...")
model.save(trainResultPath+".h5")
model.save_weights(trainResultPath +"wb.h5")

然后在pc控制台进入到工程目录：

然后在控制台输入指令：python 文件名 -predict=预测文件

完成后即可获取预测结果，我这里使用的训练图片比较少，所以预测结果可能有差异，在神经网络学习中，需要使用大量数据来进行训练。