制作网站的图片哪来,阿里云云服务器官网,做的网站,wordpress首页布局怎么修改引言#xff1a; 过去几周我一直在涉足深度学习领域#xff0c;尤其是卷积神经网络模型。最近#xff0c;谷歌围绕街景多位数字识别技术发布了一篇不错的paper。该文章描述了一个用于提取街景门牌号的单个端到端神经网络系统。然后#xff0c;作者阐述了基于同样的网络结构… 引言 过去几周我一直在涉足深度学习领域尤其是卷积神经网络模型。最近谷歌围绕街景多位数字识别技术发布了一篇不错的paper。该文章描述了一个用于提取街景门牌号的单个端到端神经网络系统。然后作者阐述了基于同样的网络结构如何来突破谷歌验证码识别系统的准确率。 为了亲身体验神经网络的实现我决定尝试设计一个可以解决类似问题的系统国内车牌号自动识别系统。设计这样一个系统的原因有3点 我应该能够参照谷歌那篇paper搭建一个同样的或者类似的网络架构谷歌提供的那个网络架构在验证码识别上相当不错那么讲道理的话用它来识别车牌号应该也会很给力。拥有一个知名的网络架构将会大大地简化我学习CNN的步骤。我可以很容易地生成训练数据车牌数。训练神经网络存在一个很大的问题就是需要大量的标签样本。通常要训练好一个网络就需要几十万张标记过的图片。 好奇心。传统的车牌号自动识别系统依赖于自己编写算法来实现车牌定位标准化分割和字符识别等功能。照这样的话实现这些系统的代码可能达到上千行。然而我比较感兴趣的是如何使用相对较少的代码和最少的专业领域知识来开发一个不错的系统。开发该项目的环境要求有Python,Tensorflow,OpenCV和NumPy等软件。源代码在这里。 首先项目结构及成果图 合成图片 genplate.py 为了训练任何一个神经网络必须提供一套拥有正确输出的训练数据。  文本和车牌背景国内正常的车牌格式但是文本颜色必须比车牌颜色更深一些。这是为了模拟真实场景的光线变化。最后再加入一些噪音这样不仅能够解释真实传感器的噪音而且能够避免过多依赖于锐化的轮廓边界而看到的将会是离焦的输入图片。font目录导入了字体包、image导入了车牌背景及噪声图片。 车牌变换采用了一种基于随机滚转、倾斜、偏转、平移以及缩放的仿射变换。每个参数允许的范围是车牌号可能被看到的所有情况的集合。比如偏转比滚转允许变化更多你更可能看到一辆汽车在拐弯而不是翻转到一边。 import PIL from PIL import ImageFont from PIL import Image from PIL import ImageDraw import cv2; import numpy as np; import os; from math import *# 生成车牌 # font ImageFont.truetype(Arial-Bold.ttf,14)index {京: 0, 沪: 1, 津: 2, 渝: 3, 冀: 4, 晋: 5, 蒙: 6, 辽: 7, 吉: 8, 黑: 9, 苏: 10, 浙: 11, 皖: 12,闽: 13, 赣: 14, 鲁: 15, 豫: 16, 鄂: 17, 湘: 18, 粤: 19, 桂: 20, 琼: 21, 川: 22, 贵: 23, 云: 24,藏: 25, 陕: 26, 甘: 27, 青: 28, 宁: 29, 新: 30, 0: 31, 1: 32, 2: 33, 3: 34, 4: 35, 5: 36,6: 37, 7: 38, 8: 39, 9: 40, A: 41, B: 42, C: 43, D: 44, E: 45, F: 46, G: 47, H: 48,J: 49, K: 50, L: 51, M: 52, N: 53, P: 54, Q: 55, R: 56, S: 57, T: 58, U: 59, V: 60,W: 61, X: 62, Y: 63, Z: 64};chars [京, 沪, 津, 渝, 冀, 晋, 蒙, 辽, 吉, 黑, 苏, 浙, 皖, 闽, 赣, 鲁, 豫, 鄂, 湘, 粤, 桂,琼, 川, 贵, 云, 藏, 陕, 甘, 青, 宁, 新, 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, A,B, C, D, E, F, G, H, J, K, L, M, N, P, Q, R, S, T, U, V, W, X,Y, Z];def AddSmudginess(img, Smu):rows r(Smu.shape[0] - 50)cols r(Smu.shape[1] - 50)adder Smu[rows:rows 50, cols:cols 50];adder cv2.resize(adder, (50, 50));# adder cv2.bitwise_not(adder)img cv2.resize(img,(50,50))img cv2.bitwise_not(img)img cv2.bitwise_and(adder, img)img cv2.bitwise_not(img)return img def rot(img,angel,shape,max_angel): 使图像轻微的畸变img 输入图像factor 畸变的参数size 为图片的目标尺寸size_o [shape[1],shape[0]]size (shape[1] int(shape[0]*cos((float(max_angel )/180) * 3.14)),shape[0])interval abs( int( sin((float(angel) /180) * 3.14)* shape[0]));pts1 np.float32([[0,0],[0,size_o[1]],[size_o[0],0],[size_o[0],size_o[1]]])if(angel0):pts2 np.float32([[interval,0],[0,size[1] ],[size[0],0 ],[size[0]-interval,size_o[1]]])else:pts2 np.float32([[0,0],[interval,size[1] ],[size[0]-interval,0 ],[size[0],size_o[1]]])M cv2.getPerspectiveTransform(pts1,pts2);dst cv2.warpPerspective(img,M,size);return dst; def rotRandrom(img, factor, size):shape size;pts1 np.float32([[0, 0], [0, shape[0]], [shape[1], 0], [shape[1], shape[0]]])pts2 np.float32([[r(factor), r(factor)], [ r(factor), shape[0] - r(factor)], [shape[1] - r(factor), r(factor)],[shape[1] - r(factor), shape[0] - r(factor)]])M cv2.getPerspectiveTransform(pts1, pts2);dst cv2.warpPerspective(img, M, size);return dst; def tfactor(img):hsv cv2.cvtColor(img,cv2.COLOR_BGR2HSV);hsv[:,:,0] hsv[:,:,0]*(0.8 np.random.random()*0.2);hsv[:,:,1] hsv[:,:,1]*(0.3 np.random.random()*0.7);hsv[:,:,2] hsv[:,:,2]*(0.2 np.random.random()*0.8);img cv2.cvtColor(hsv,cv2.COLOR_HSV2BGR);return img def random_envirment(img,data_set):indexr(len(data_set))env cv2.imread(data_set[index])env cv2.resize(env,(img.shape[1],img.shape[0]))bak (img0);bak bak.astype(np.uint8)*255;inv cv2.bitwise_and(bak,env)img cv2.bitwise_or(inv,img)return img def GenCh(f,val):imgImage.new(RGB, (45,70),(255,255,255))draw ImageDraw.Draw(img)draw.text((0, 3),val,(0,0,0),fontf)img img.resize((23,70))A np.array(img)return A def GenCh1(f,val):imgImage.new(RGB, (23,70),(255,255,255))draw ImageDraw.Draw(img)#draw.text((0, 2),val.decode(utf-8),(0,0,0),fontf)draw.text((0, 2),val,(0,0,0),fontf)A np.array(img)return A def AddGauss(img, level):return cv2.blur(img, (level * 2 1, level * 2 1));def r(val):return int(np.random.random() * val)def AddNoiseSingleChannel(single):diff 255-single.max();noise np.random.normal(0,1r(6),single.shape);noise (noise - noise.min())/(noise.max()-noise.min())noise diff*noise;noise noise.astype(np.uint8)dst single noisereturn dstdef addNoise(img,sdev 0.5,avg10):img[:,:,0] AddNoiseSingleChannel(img[:,:,0]);img[:,:,1] AddNoiseSingleChannel(img[:,:,1]);img[:,:,2] AddNoiseSingleChannel(img[:,:,2]);return img;class GenPlate:def __init__(self,fontCh,fontEng,NoPlates):self.fontC ImageFont.truetype(fontCh,43,0);self.fontE ImageFont.truetype(fontEng,60,0);self.imgnp.array(Image.new(RGB, (226,70),(255,255,255)))self.bg cv2.resize(cv2.imread(./images/template.bmp),(226,70));self.smu cv2.imread(./images/smu2.jpg);self.noplates_path [];for parent,parent_folder,filenames in os.walk(NoPlates):for filename in filenames:path parent/filename;self.noplates_path.append(path);def draw(self,val):offset 2 ;self.img[0:70,offset8:offset823] GenCh(self.fontC,val[0]);self.img[0:70,offset8236:offset823623] GenCh1(self.fontE,val[1]);for i in range(5):base offset82362317 i*23 i*6 ;self.img[0:70, base : base23] GenCh1(self.fontE,val[i2]);return self.imgdef generate(self,text):if len(text) 7:fg self.draw(text.encode(utf-8).decode(encodingutf-8));fg cv2.bitwise_not(fg);com cv2.bitwise_or(fg,self.bg);com rot(com,r(60)-30,com.shape,30);com rotRandrom(com,10,(com.shape[1],com.shape[0]));#com AddSmudginess(com,self.smu)com tfactor(com)com random_envirment(com,self.noplates_path);com AddGauss(com, 1r(4));com addNoise(com);return comdef genPlateString(self,pos,val):plateStr ;box [0,0,0,0,0,0,0];if(pos!-1):box[pos]1;for unit,cpos in zip(box,range(len(box))):if unit 1:plateStr valelse:if cpos 0:plateStr chars[r(31)]elif cpos 1:plateStr chars[41r(24)]else:plateStr chars[31 r(34)]return plateStr;def genBatch(self, batchSize,pos,charRange, outputPath,size):if (not os.path.exists(outputPath)):os.mkdir(outputPath)for i in range(batchSize):plateStr G.genPlateString(-1,-1)img G.generate(plateStr);img cv2.resize(img,size);cv2.imwrite(outputPath / str(i).zfill(2) .jpg, img);# return img G GenPlate(./font/platech.ttf,./font/platechar.ttf,./NoPlates) G.genBatch(15,2,range(31,65),./plate,(272,72)) #注释原因为每次其他模块运行若导入该库都会刷性该函数 inputdata.py 省略.... 产生用于训练的数据 # 批量生成车牌供算法调用 网络结构  # 算法的核心卷积神经网络 使用的卷积神经网络结构如model.py所示 import tensorflow as tf import numpy as np# 算法的核心卷积神经网络 def inference (images,keep_prob):Build the modelArgs:image: image batch,4D tensor,tf.float32,[batch_size,height,width,channels]Returns:output tensor with the computed logits,float,[batch_size,65]# conv1with tf.variable_scope(conv1) as scope:weights tf.get_variable(weights,shape [3,3,3,32],dtype tf.float32,initializertf.truncated_normal_initializer(stddev0.1,dtypetf.float32))conv tf.nn.conv2d(images,weights,strides[1,1,1,1],paddingVALID)biases tf.get_variable(biases,shape[32],dtypetf.float32,initializertf.constant_initializer(0.1))pre_activation tf.nn.bias_add(conv,biases)conv1 tf.nn.relu(pre_activation,name scope.name)# conv2with tf.variable_scope(conv2) as scope:weights tf.get_variable(weights,shape[3,3,32,32],dtypetf.float32,initializertf.truncated_normal_initializer(stddev0.1,dtypetf.float32))conv tf.nn.conv2d(conv1,weights,strides[1,1,1,1],paddingVALID)biases tf.get_variable(biases,shape[32],dtypetf.float32,initializertf.constant_initializer(0.1))pre_activation tf.nn.bias_add(conv,biases)conv2 tf.nn.relu(pre_activation,name scope.name)with tf.variable_scope(max_pooling1) as scope:pool1 tf.nn.max_pool(conv2,ksize [1,2,2,1],strides [1,2,2,1],paddingVALID,namepooling1)#conv3with tf.variable_scope(conv3) as scope:weights tf.get_variable(weights,shape[3,3,32,64],dtypetf.float32,initializertf.truncated_normal_initializer(stddev0.1,dtypetf.float32))conv tf.nn.conv2d(pool1,weights,strides[1,1,1,1],paddingVALID)biases tf.get_variable(biases,shape[64],dtype tf.float32,initializer tf.constant_initializer(0.1))pre_activation tf.nn.bias_add(conv,biases)conv3 tf.nn.relu(pre_activation,namescope.name)#conv4with tf.variable_scope(conv4) as scope:weights tf.get_variable(weights,shape[3,3,64,64],dtypetf.float32,initializertf.truncated_normal_initializer(stddev0.1,dtypetf.float32))conv tf.nn.conv2d(conv3,weights,strides[1,1,1,1],paddingVALID)biases tf.get_variable(biases,shape[64],dtypetf.float32,initializertf.constant_initializer(0.1))pre_activation tf.nn.bias_add(conv,biases)conv4 tf.nn.relu(pre_activation,namescope.name)with tf.variable_scope(max_pooling2) as scope:pool2 tf.nn.max_pool(conv4,ksize[1,2,2,1],strides[1,2,2,1],paddingVALID,namepooling2)#conv5with tf.variable_scope(conv5) as scope:weights tf.get_variable(weights,shape[3,3,64,128],dtypetf.float32,initializertf.truncated_normal_initializer(stddev0.1,dtypetf.float32))conv tf.nn.conv2d(pool2,weights,strides[1,1,1,1],paddingVALID)biases tf.get_variable(biases,shape[128],dtypetf.float32,initializertf.constant_initializer(0.1))pre_activation tf.nn.bias_add(conv,biases)conv5 tf.nn.relu(pre_activation,namescope.name)#conv6with tf.variable_scope(conv6) as scope:weights tf.get_variable(weights,shape[3,3,128,128],dtypetf.float32,initializertf.truncated_normal_initializer(stddev0.1,dtypetf.float32))conv tf.nn.conv2d(conv5,weights,strides[1,1,1,1],paddingVALID)biases tf.get_variable(biases,shape[128],dtypetf.float32,initializertf.constant_initializer(0.1))pre_activation tf.nn.bias_add(conv,biases)conv6 tf.nn.relu(pre_activation,namescope.name)#pool3with tf.variable_scope(max_pool3) as scope:pool3 tf.nn.max_pool(conv6,ksize[1,2,2,1],strides[1,2,2,1],paddingVALID,namepool3)#%%#fc1_flattenwith tf.variable_scope(fc1) as scope:shp pool3.get_shape()flattened_shape shp[1].value*shp[2].value*shp[3].valuereshape tf.reshape(pool3,[-1,flattened_shape])fc1 tf.nn.dropout(reshape,keep_prob,namefc1_dropdot)# 全连接层# 第一个全连接层识别车牌上的省with tf.variable_scope(fc21) as scope:weights tf.get_variable(weights,shape[flattened_shape,65],dtypetf.float32,initializertf.truncated_normal_initializer(stddev0.005,dtypetf.float32))biases tf.get_variable(biases,shape[65],dtypetf.float32,initializer tf.truncated_normal_initializer(0.1))fc21 tf.matmul(fc1,weights)biases# 第二个全连接层识别车牌上的市with tf.variable_scope(fc22) as scope:weights tf.get_variable(weights,shape[flattened_shape,65],dtypetf.float32,initializertf.truncated_normal_initializer(stddev0.005,dtypetf.float32))biases tf.get_variable(biases,shape[65],dtypetf.float32,initializer tf.truncated_normal_initializer(0.1))fc22 tf.matmul(fc1,weights)biases# 第三个全连接层识别车牌第三位的字母或者数字with tf.variable_scope(fc23) as scope:weights tf.get_variable(weights,shape[flattened_shape,65],dtypetf.float32,initializertf.truncated_normal_initializer(stddev0.005,dtypetf.float32))biases tf.get_variable(biases,shape[65],dtypetf.float32,initializer tf.truncated_normal_initializer(0.1))fc23 tf.matmul(fc1,weights)biaseswith tf.variable_scope(fc24) as scope:weights tf.get_variable(weights,shape[flattened_shape,65],dtypetf.float32,initializertf.truncated_normal_initializer(stddev0.005,dtypetf.float32))biases tf.get_variable(biases,shape[65],dtypetf.float32,initializer tf.truncated_normal_initializer(0.1))fc24 tf.matmul(fc1,weights)biaseswith tf.variable_scope(fc25) as scope:weights tf.get_variable(weights,shape[flattened_shape,65],dtypetf.float32,initializertf.truncated_normal_initializer(stddev0.005,dtypetf.float32))biases tf.get_variable(biases,shape[65],dtypetf.float32,initializer tf.truncated_normal_initializer(0.1))fc25 tf.matmul(fc1,weights)biaseswith tf.variable_scope(fc26) as scope:weights tf.get_variable(weights,shape[flattened_shape,65],dtypetf.float32,initializertf.truncated_normal_initializer(stddev0.005,dtypetf.float32))biases tf.get_variable(biases,shape[65],dtypetf.float32,initializer tf.truncated_normal_initializer(0.1))fc26 tf.matmul(fc1,weights)biaseswith tf.variable_scope(fc27) as scope:weights tf.get_variable(weights,shape[flattened_shape,65],dtypetf.float32,initializertf.truncated_normal_initializer(stddev0.005,dtypetf.float32))biases tf.get_variable(biases,shape[65],dtypetf.float32,initializer tf.truncated_normal_initializer(0.1))fc27 tf.matmul(fc1,weights)biasesreturn fc21,fc22,fc23,fc24,fc25,fc26,fc27 #shape [7,batch_size,65]# 卷积神经网络返回的输出层进行交叉熵计算 def losses(logits1,logits2,logits3,logits4,logits5,logits6,logits7,labels):Compute loss from logits and labelsArgs:logits: logits tensor, float, [7*batch_size, 65]labels: label tensor, tf.int32, [7*batch_size]Returns:loss tensor of float typelabels tf.convert_to_tensor(labels,tf.int32)with tf.variable_scope(loss1) as scope:cross_entropy tf.nn.sparse_softmax_cross_entropy_with_logits(logitslogits1, labelslabels[:,0], namexentropy_per_example)#cross_entropy tf.nn.softmax_cross_entropy_with_logits(logitslogits,labelslabels,namexentropy_per_example)loss1 tf.reduce_mean(cross_entropy, nameloss1)tf.summary.scalar(scope.name/loss1, loss1)with tf.variable_scope(loss2) as scope:cross_entropy tf.nn.sparse_softmax_cross_entropy_with_logits(logitslogits2, labelslabels[:,1], namexentropy_per_example)#cross_entropy tf.nn.softmax_cross_entropy_with_logits(logitslogits,labelslabels,namexentropy_per_example)loss2 tf.reduce_mean(cross_entropy, nameloss2)tf.summary.scalar(scope.name/loss2, loss2)with tf.variable_scope(loss3) as scope:cross_entropy tf.nn.sparse_softmax_cross_entropy_with_logits(logitslogits3, labelslabels[:,2], namexentropy_per_example)#cross_entropy tf.nn.softmax_cross_entropy_with_logits(logitslogits,labelslabels,namexentropy_per_example)loss3 tf.reduce_mean(cross_entropy, nameloss3)tf.summary.scalar(scope.name/loss3, loss3)with tf.variable_scope(loss4) as scope:cross_entropy tf.nn.sparse_softmax_cross_entropy_with_logits(logitslogits4, labelslabels[:,3], namexentropy_per_example)#cross_entropy tf.nn.softmax_cross_entropy_with_logits(logitslogits,labelslabels,namexentropy_per_example)loss4 tf.reduce_mean(cross_entropy, nameloss4)tf.summary.scalar(scope.name/loss4, loss4)with tf.variable_scope(loss5) as scope:cross_entropy tf.nn.sparse_softmax_cross_entropy_with_logits(logitslogits5, labelslabels[:,4], namexentropy_per_example)#cross_entropy tf.nn.softmax_cross_entropy_with_logits(logitslogits,labelslabels,namexentropy_per_example)loss5 tf.reduce_mean(cross_entropy, nameloss5)tf.summary.scalar(scope.name/loss5, loss5)with tf.variable_scope(loss6) as scope:cross_entropy tf.nn.sparse_softmax_cross_entropy_with_logits(logitslogits6, labelslabels[:,5], namexentropy_per_example)#cross_entropy tf.nn.softmax_cross_entropy_with_logits(logitslogits,labelslabels,namexentropy_per_example)loss6 tf.reduce_mean(cross_entropy, nameloss6)tf.summary.scalar(scope.name/loss6, loss6)with tf.variable_scope(loss7) as scope:cross_entropy tf.nn.sparse_softmax_cross_entropy_with_logits(logitslogits7, labelslabels[:,6], namexentropy_per_example)#cross_entropy tf.nn.softmax_cross_entropy_with_logits(logitslogits,labelslabels,namexentropy_per_example)loss7 tf.reduce_mean(cross_entropy, nameloss7)tf.summary.scalar(scope.name/loss7, loss7)return loss1,loss2,loss3,loss4,loss5,loss6,loss7# 最优化自适应梯度下降进行优化 def trainning( loss1,loss2,loss3,loss4,loss5,loss6,loss7, learning_rate):Training ops, the Op returned by this function is what must be passed tosess.run() call to cause the model to train.Args:loss: loss tensor, from losses()Returns:train_op: The op for trainningwith tf.name_scope(optimizer1):optimizer1 tf.train.AdamOptimizer(learning_rate learning_rate)global_step tf.Variable(0, nameglobal_step, trainableFalse)train_op1 optimizer1.minimize(loss1, global_step global_step)with tf.name_scope(optimizer2):optimizer2 tf.train.AdamOptimizer(learning_rate learning_rate)global_step tf.Variable(0, nameglobal_step, trainableFalse)train_op2 optimizer2.minimize(loss2, global_step global_step)with tf.name_scope(optimizer3):optimizer3 tf.train.AdamOptimizer(learning_rate learning_rate)global_step tf.Variable(0, nameglobal_step, trainableFalse)train_op3 optimizer3.minimize(loss3, global_step global_step)with tf.name_scope(optimizer4):optimizer4 tf.train.AdamOptimizer(learning_rate learning_rate)global_step tf.Variable(0, nameglobal_step, trainableFalse)train_op4 optimizer4.minimize(loss4, global_step global_step)with tf.name_scope(optimizer5):optimizer5 tf.train.AdamOptimizer(learning_rate learning_rate)global_step tf.Variable(0, nameglobal_step, trainableFalse)train_op5 optimizer5.minimize(loss5, global_step global_step)with tf.name_scope(optimizer6):optimizer6 tf.train.AdamOptimizer(learning_rate learning_rate)global_step tf.Variable(0, nameglobal_step, trainableFalse)train_op6 optimizer6.minimize(loss6, global_step global_step)with tf.name_scope(optimizer7):optimizer7 tf.train.AdamOptimizer(learning_rate learning_rate)global_step tf.Variable(0, nameglobal_step, trainableFalse)train_op7 optimizer7.minimize(loss7, global_step global_step)return train_op1,train_op2,train_op3,train_op4,train_op5,train_op6,train_op7# 对模型评估 def evaluation(logits1,logits2,logits3,logits4,logits5,logits6,logits7,labels):Evaluate the quality of the logits at predicting the label.Args:logits: Logits tensor, float - [batch_size, NUM_CLASSES].labels: Labels tensor, int32 - [batch_size], with values in therange [0, NUM_CLASSES).Returns:A scalar int32 tensor with the number of examples (out of batch_size)that were predicted correctly.logits_all tf.concat([logits1,logits2,logits3,logits4,logits5,logits6,logits7],0)labels tf.convert_to_tensor(labels,tf.int32)labels_all tf.reshape(tf.transpose(labels),[-1])with tf.variable_scope(accuracy) as scope:correct tf.nn.in_top_k(logits_all, labels_all, 1)correct tf.cast(correct, tf.float16)accuracy tf.reduce_mean(correct)tf.summary.scalar(scope.name/accuracy, accuracy)return accuracy 训练模型train.py import os import numpy as np import tensorflow as tf from input_data import OCRIter import model import time import datetime# train训练 img_w 272 img_h 72 num_label7 batch_size 8 # count 30000 count 500000 learning_rate 0.0001#默认参数[N,H,W,C] image_holder tf.placeholder(tf.float32,[batch_size,img_h,img_w,3]) label_holder tf.placeholder(tf.int32,[batch_size,7]) keep_prob tf.placeholder(tf.float32)logs_train_dir rH:\GPpython\第三阶段数据分析\神经网络\Licence_plate_recognize\Licence_plate_recognize\train_resultdef get_batch():data_batch OCRIter(batch_size,img_h,img_w)image_batch,label_batch data_batch.iter()image_batch1 np.array(image_batch)label_batch1 np.array(label_batch)return image_batch1,label_batch1def fit():train_logits1, train_logits2, train_logits3, train_logits4, train_logits5, train_logits6, train_logits7 model.inference(image_holder, keep_prob)train_loss1, train_loss2, train_loss3, train_loss4, train_loss5, train_loss6, train_loss7 model.losses(train_logits1, train_logits2, train_logits3, train_logits4, train_logits5, train_logits6, train_logits7,label_holder)train_op1, train_op2, train_op3, train_op4, train_op5, train_op6, train_op7 model.trainning(train_loss1,train_loss2,train_loss3,train_loss4,train_loss5,train_loss6,train_loss7,learning_rate)train_acc model.evaluation(train_logits1, train_logits2, train_logits3, train_logits4, train_logits5,train_logits6, train_logits7, label_holder)input_image tf.summary.image(input, image_holder)summary_op tf.summary.merge(tf.get_collection(tf.GraphKeys.SUMMARIES))sess tf.Session()train_writer tf.summary.FileWriter(logs_train_dir, sess.graph)saver tf.train.Saver()sess.run(tf.global_variables_initializer())start_time1 time.time()for step in range(count):x_batch, y_batch get_batch()start_time2 time.time()time_str datetime.datetime.now().isoformat()feed_dict {image_holder: x_batch, label_holder: y_batch, keep_prob: 0.5}_, _, _, _, _, _, _, tra_loss1, tra_loss2, tra_loss3, tra_loss4, tra_loss5, tra_loss6, tra_loss7, acc, summary_str sess.run([train_op1, train_op2, train_op3, train_op4, train_op5, train_op6, train_op7, train_loss1, train_loss2,train_loss3, train_loss4, train_loss5, train_loss6, train_loss7, train_acc, summary_op], feed_dict)train_writer.add_summary(summary_str, step)duration time.time() - start_time2tra_all_loss tra_loss1 tra_loss2 tra_loss3 tra_loss4 tra_loss5 tra_loss6 tra_loss7# print(y_batch) #仅测试代码训练实际样本与标签是否一致if step % 10 0:sec_per_batch float(duration)print(%s : Step %d,train_loss %.2f,acc %.2f,sec/batch%.3f % (time_str, step, tra_all_loss, acc, sec_per_batch))if step % 10000 0 or (step 1) count:checkpoint_path os.path.join(logs_train_dir, model.ckpt)saver.save(sess, checkpoint_path, global_stepstep)sess.close()print(time.time() - start_time1) if __name__ __main__:# fit()data,labels get_batch()print(data.shape)print(labels.shape)print(labels[0]) 除了输出层使用ReLU激活函数之外所有层都采用深度神经网络的标准结构。指示存在的节点使用sigmoid激活函数典型地用于二值输出。其他输出节点使用softmax贯穿字符结果是每一列的概率之和为1是模型化离散概率分布的标准方法。 根据标签和网络输出的交叉熵来定义损失函数。为了数值稳定性利用softmax_cross_entropy_with_logits和sigmoid_cross_entropy_with_logits将最后一层的激活函数卷入交叉熵的计算。关于对交叉熵详细而直观的介绍可以参考Michael A. Nielsen的free online book中查看这一节。 使用一块nVidia GTX 960m花费大约数小时来训练train.py通过CPU的一个后台进程来运行训练数据的生成。 输出处理 import tensorflow as tf import numpy as np import os from PIL import Image import cv2 import matplotlib.pyplot as plt import model import genplate index {京: 0, 沪: 1, 津: 2, 渝: 3, 冀: 4, 晋: 5, 蒙: 6, 辽: 7, 吉: 8, 黑: 9, 苏: 10, 浙: 11, 皖: 12,闽: 13, 赣: 14, 鲁: 15, 豫: 16, 鄂: 17, 湘: 18, 粤: 19, 桂: 20, 琼: 21, 川: 22, 贵: 23, 云: 24,藏: 25, 陕: 26, 甘: 27, 青: 28, 宁: 29, 新: 30, 0: 31, 1: 32, 2: 33, 3: 34, 4: 35, 5: 36,6: 37, 7: 38, 8: 39, 9: 40, A: 41, B: 42, C: 43, D: 44, E: 45, F: 46, G: 47, H: 48,J: 49, K: 50, L: 51, M: 52, N: 53, P: 54, Q: 55, R: 56, S: 57, T: 58, U: 59, V: 60,W: 61, X: 62, Y: 63, Z: 64};chars [京, 沪, 津, 渝, 冀, 晋, 蒙, 辽, 吉, 黑, 苏, 浙, 皖, 闽, 赣, 鲁, 豫, 鄂, 湘, 粤, 桂,琼, 川, 贵, 云, 藏, 陕, 甘, 青, 宁, 新, 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, A,B, C, D, E, F, G, H, J, K, L, M, N, P, Q, R, S, T, U, V, W, X,Y, Z];Test one image against the saved models and parametersglobal pic def get_one_image(test):Randomly pick one image from training dataReturn: ndarryG genplate.GenPlate(./font/platech.ttf, ./font/platechar.ttf, ./NoPlates)G.genBatch(15, 2, range(31, 65), ./plate, (272, 72)) # 注释原因为每次其他模块运行若导入该库都会刷性该函数n len(test)ind np.random.randint(0,n)img_dir test[ind]image_show Image.open(img_dir)plt.imshow(image_show)#image image.resize([120,30])image cv2.imread(img_dir)global picpic image# cv2.imshow(image, image)# cv2.waitKey(0)img np.multiply(image,1/255.0)#image np.array(img)#image img.transpose(1,0,2)image np.array([img])print(image.shape)return imagebatch_size 1 x tf.placeholder(tf.float32,[batch_size,72,272,3]) keep_prob tf.placeholder(tf.float32)test_dir rH:/GPpython/Licence_plate_recognize/plate/ test_image [] for file in os.listdir(test_dir):test_image.append(test_dir file) test_image list(test_image)image_array get_one_image(test_image)#logit model.inference(x,keep_prob) logit1,logit2,logit3,logit4,logit5,logit6,logit7 model.inference(x,keep_prob)#logit1 tf.nn.softmax(logit1) #logit2 tf.nn.softmax(logit2) #logit3 tf.nn.softmax(logit3) #logit4 tf.nn.softmax(logit4) #logit5 tf.nn.softmax(logit5) #logit6 tf.nn.softmax(logit6) #logit7 tf.nn.softmax(logit7)logs_train_dir rH:/GPpython/Licence_plate_recognize/train_result/ saver tf.train.Saver()with tf.Session() as sess:print (Reading checkpoint...)ckpt tf.train.get_checkpoint_state(logs_train_dir)if ckpt and ckpt.model_checkpoint_path:global_step ckpt.model_checkpoint_path.split(/)[-1].split(-)[-1]saver.restore(sess, ckpt.model_checkpoint_path)print(Loading success, global_step is %s % global_step)else:print(No checkpoint file found)pre1,pre2,pre3,pre4,pre5,pre6,pre7 sess.run([logit1,logit2,logit3,logit4,logit5,logit6,logit7], feed_dict{x: image_array,keep_prob:1.0})prediction np.reshape(np.array([pre1,pre2,pre3,pre4,pre5,pre6,pre7]),[-1,65])#prediction np.array([[pre1],[pre2],[pre3],[pre4],[pre5],[pre6],[pre7]])#print(prediction)max_index np.argmax(prediction,axis1)print(max_index)line for i in range(prediction.shape[0]):if i 0:result np.argmax(prediction[i][0:31])if i 1:result np.argmax(prediction[i][41:65])41if i 1:result np.argmax(prediction[i][31:65])31line chars[result] print (predicted: line)cv2.imshow(pic,pic) cv2.waitKeyEx(0) 总结 我已经开源了一个拥有相对较短代码系统它不用导入任何特定领域的库以及不需要太多特定领域的知识就能够实现车牌号自动识别。此外我还通过在线合成图片的方法解决了上千张训练图片的需求问题通常是在深度神经网络的情况下。 另一方面我的系统也存在一些缺点 只适用于特定车牌号。尤其是网络结构明确假定了输出只有7个字符。只适用于特定字体。速度太慢。该系统运行一张适当尺寸的图片要花费几秒钟。 为了解决第1个问题谷歌团队将他们的网络结构的高层拆分成了多个子网络每一个子网络用于假定输出号码中的不同号码位。还有一个并行的子网络来决定存在多少号码。我觉得这种方法可以应用到这儿但是我没有在这个项目中实现。 关于第2点我在上面举过例子由于字体的稍微不同只能使用于国内车牌。如果尝试着检测US车牌号的话误检将会更加严重因为US车牌号字体类型更多。一个可能的解决方案就是使得训练数据有更多不同的字体类型可选择尽管还不清楚需要多少字体类型才能成功。 第3点提到的速度慢的问题是扼杀许多应用的cancer在一个相当强大的GPU上处理一张适当尺寸的输入图片就要花费几秒钟。我认为不引进一种级联式结构的检测网络就想避开这个问题是不太可能的比如Haar级联HOG检测器或者一个更简单的神经网络。 我很有兴趣去尝试和其他机器学习方法的比较会怎样特别是姿态回归看起来有希望最后可能会附加一个最基本的分类阶段。如果使用了像scikit-learn这样的机器学习库那么应该同样简单。