機(jī)器學(xué)習(xí)練習(xí)目錄 一、理解人臉圖像特征提取的各種方法的特征 1.HOG 2.Dlib 3.卷積神經(jīng)網(wǎng)絡(luò)（CNN） 二、卷積神經(jīng)網(wǎng)絡(luò)（CNN）笑臉數(shù)據(jù)集（genki4k）正負(fù)樣本的劃分、模型訓(xùn)練和測試的過程，輸出模型訓(xùn)練精度和測試精度，完成對人臉微笑與否的模型訓(xùn)練 1.查看運(yùn)行的環(huán)境 2.數(shù)據(jù)集準(zhǔn)備 3.網(wǎng)絡(luò)模型 4.資料預(yù)處理 5.開始訓(xùn)練 6.使用數(shù)據(jù)填充 7.對人臉微笑與否的模型訓(xùn)練 三、卷積神經(jīng)網(wǎng)絡(luò)（CNN）對口罩?jǐn)?shù)據(jù)集正負(fù)樣本的劃分、模型訓(xùn)練和測試的過程，輸出模型訓(xùn)練精度和測試精度，完成對口罩佩戴與否的模型訓(xùn)練 1.數(shù)據(jù)集準(zhǔn)備 2.網(wǎng)絡(luò)模型 3.資料預(yù)處理 4.開始訓(xùn)練 5.使用數(shù)據(jù)填充 6.對人臉戴口罩與否的模型訓(xùn)練 四、完成一個(gè)攝像頭采集自己人臉、并對表情（笑臉/非笑臉、戴口罩和沒戴口罩）的實(shí)時(shí)分類判讀（輸出分類文字）的程序 1.笑臉/非笑臉實(shí)時(shí)分類判讀（輸出分類文字）的程序 2.戴口罩和沒戴口罩的實(shí)時(shí)分類判讀（輸出分類文字）的程序
一、理解人臉圖像特征提取的各種方法的特征
1.HOG
介紹： 方向梯度直方圖（Histogram of Oriented Gradient, HOG）特征是一種在計(jì)算機(jī)視覺和圖像處理中用來進(jìn)行物體檢測的特征描述子。它通過計(jì)算和統(tǒng)計(jì)圖像局部區(qū)域的梯度方向直方圖來構(gòu)成特征。 主要思想： 在一副圖像中，局部目標(biāo)的表象和形狀能夠被梯度或邊緣的方向密度分布很好地描述。（本質(zhì)：梯度的統(tǒng)計(jì)信息，而梯度主要存在于邊緣的地方）。 具體的實(shí)現(xiàn)方法： 首先將圖像分成小的連通區(qū)域，把它叫細(xì)胞單元。然后采集細(xì)胞單元中各像素點(diǎn)的梯度的或邊緣的方向直方圖。最后把這些直方圖組合起來就可以構(gòu)成特征描述器。 步驟： 灰度圖像轉(zhuǎn)換、梯度計(jì)算、分網(wǎng)格的梯度方向直方圖、快描述子、快描述子歸一化、特征數(shù)據(jù)與檢測窗口、匹配方法。 優(yōu)點(diǎn)： ① HOG是在圖像的局部方格單元上操作，它對圖像幾何的和光學(xué)的形變都能保持很好的不變性，這兩種形變只會出現(xiàn)在更大的空間領(lǐng)域上；②在粗的空域抽樣、精細(xì)的方向抽樣以及較強(qiáng)的局部光學(xué)歸一化等條件下，只要行人大體上能夠保持直立的姿勢，可以容許行人有一些細(xì)微的肢體動(dòng)作，這些細(xì)微的動(dòng)作可以被忽略而不影響檢測效果。 綜上所述，HOG特征是特別適合于做圖像中的人體檢測的。 參考鏈接：https://zhuanlan.zhihu.com/p/104670289
2.Dlib
介紹： Dlib是一款優(yōu)秀的跨平臺開源的C++工具庫，該庫使用C++編寫，具有優(yōu)異的性能。Dlib庫提供的功能十分豐富，包括線性代數(shù)，圖像處理，機(jī)器學(xué)習(xí)，網(wǎng)絡(luò)，最優(yōu)化算法等眾多功能。同時(shí)該庫也提供了Python接口。 核心原理： 使用了圖像Hog特征來表示人臉，和其他特征提取算子相比，它對圖像的幾何和光學(xué)的形變都能保持很好的不變形。該特征與LBP特征，Harr特征共同作為三種經(jīng)典的圖像特征，該特征提取算子通常和支持向量機(jī)(SVM)算法搭配使用，用在物體檢測場景。 該算法大致思路： 對正樣本(即包含人臉的圖像)數(shù)據(jù)集提取Hog特征，得到Hog特征描述子。 對負(fù)樣本(即不包含人臉的圖像)數(shù)據(jù)集提取Hog特征，得到Hog描述子。 其中負(fù)樣本數(shù)據(jù)集中的數(shù)據(jù)量要遠(yuǎn)遠(yuǎn)大于正樣本數(shù)據(jù)集中的樣本數(shù)，負(fù)樣本圖像可以使用不含人臉的圖片進(jìn)行隨機(jī)裁剪獲取。 利用支持向量機(jī)算法訓(xùn)練正負(fù)樣本，顯然這是一個(gè)二分類問題，可以得到訓(xùn)練后的模型。 利用該模型進(jìn)行負(fù)樣本難例檢測，也就是難分樣本挖掘(hard-negtive mining)，以便提高最終模型的分類能力。具體思路為：對訓(xùn)練集里的負(fù)樣本不斷進(jìn)行縮放，直至與模板匹配位置，通過模板滑動(dòng)串口搜索匹配（該過程即多尺度檢測過程），如果分類器誤檢出非人臉區(qū)域則截取該部分圖像加入到負(fù)樣本中。 集合難例樣本重新訓(xùn)練模型，反復(fù)如此得到最終分類模型。 應(yīng)用最終訓(xùn)練出的分類器檢測人臉圖片，對該圖片的不同尺寸進(jìn)行滑動(dòng)掃描，提取Hog特征，并用分類器分類。如果檢測判定為人臉，則將其標(biāo)定出來，經(jīng)過一輪滑動(dòng)掃描后必然會出現(xiàn)同一個(gè)人臉被多次標(biāo)定的情況，這就用NMS完成收尾工作即可。 參考鏈接：https://zhuanlan.zhihu.com/p/92132280
3.卷積神經(jīng)網(wǎng)絡(luò)（CNN）
介紹： 卷積神經(jīng)網(wǎng)絡(luò)（Convolutional Neural Networks, CNN）是一類包含卷積計(jì)算且具有深度結(jié)構(gòu)的前饋神經(jīng)網(wǎng)絡(luò)，是深度學(xué)習(xí)的代表算法之一 。卷積神經(jīng)網(wǎng)絡(luò)具有表征學(xué)習(xí)能力，能夠按其階層結(jié)構(gòu)對輸入信息進(jìn)行平移不變分類，因此也被稱為“平移不變?nèi)斯ど窠?jīng)網(wǎng)絡(luò)（SIANN）"。 CNN原理： 卷積神經(jīng)網(wǎng)絡(luò)（CNN）主要是用于圖像識別領(lǐng)域，它指的是一類網(wǎng)絡(luò)，而不是某一種，其包含很多不同種結(jié)構(gòu)的網(wǎng)絡(luò)。不同的網(wǎng)絡(luò)結(jié)構(gòu)通常表現(xiàn)會不一樣。 所有CNN最終都是把一張圖片轉(zhuǎn)化為特征向量。就像上圖VGG網(wǎng)絡(luò)一樣，通過多層的卷積，池化，全連接，降低圖片維度，最后轉(zhuǎn)化成了一個(gè)一維向量。這個(gè)向量就包含了圖片的特征，當(dāng)然這個(gè)特征不是肉眼上的圖片特征，而是針對于神經(jīng)網(wǎng)絡(luò)的特征。 卷積神經(jīng)網(wǎng)絡(luò)的前向傳播過程 前向傳播中的卷積操作 前向傳播中的池化操作 前向傳播中的全連接 卷積神經(jīng)網(wǎng)絡(luò)的反向傳播過程 卷積神經(jīng)網(wǎng)絡(luò)的權(quán)值更新 卷積神經(jīng)網(wǎng)絡(luò)的訓(xùn)練過程流程圖： 圖片分類 相似圖搜索 對抗樣本（比如輸入一張貓的圖片，人為的修改一點(diǎn)圖片數(shù)據(jù)，肉眼上看還是一只貓，但是你告訴神經(jīng)網(wǎng)絡(luò)這是狗。） 參考鏈接：https://zhuanlan.zhihu.com/p/95158245
二、卷積神經(jīng)網(wǎng)絡(luò)（CNN）笑臉數(shù)據(jù)集（genki4k）正負(fù)樣本的劃分、模型訓(xùn)練和測試的過程，輸出模型訓(xùn)練精度和測試精度，完成對人臉微笑與否的模型訓(xùn)練
1.查看運(yùn)行的環(huán)境
import platform
import tensorflow
import keras
print ( "Platform: {}" . format ( platform
. platform
( ) ) )
print ( "Tensorflow version: {}" . format ( tensorflow
. __version__
) )
print ( "Keras version: {}" . format ( keras
. __version__
) )
Platform: Windows-7-6.1.7601-SP1
Tensorflow version: 1.2.1
Keras version: 2.1.2
2.數(shù)據(jù)集準(zhǔn)備
下載圖像數(shù)據(jù)集genki4k.tar，把它解壓到相應(yīng)的目錄（我放在了D:\mango目錄下）， 解壓出來看見的原始數(shù)據(jù)集內(nèi)容如圖所示 手動(dòng)裁剪過程 其中裁剪之前的圖片 對圖像進(jìn)行了裁剪，提取大頭照，并把一些不符合的圖片進(jìn)行了剔除，執(zhí)行完裁剪后的顯示效果 由于各種原因沒有裁剪剔除的圖片 導(dǎo)入需要的包 import keras
import matplotlib
. pyplot
as plt
import matplotlib
. image
as mpimg
import numpy
as np
from IPython
. display
import Image
import os
劃分?jǐn)?shù)據(jù)集（在當(dāng)前寫代碼的同級目錄下會產(chǎn)生一個(gè)mangoout的文件夾，包括 train :訓(xùn)練集；validation：驗(yàn)證集；test：測試集。）
original_dataset_dir
= 'D:\\mango\\files\\cutmango'
base_dir
= 'mangoout'
os
. mkdir
( base_dir
)
train_dir
= os
. path
. join
( base_dir
, 'train' )
os
. mkdir
( train_dir
)
validation_dir
= os
. path
. join
( base_dir
, 'validation' )
os
. mkdir
( validation_dir
)
test_dir
= os
. path
. join
( base_dir
, 'test' )
os
. mkdir
( test_dir
)
train_smile_dir
= os
. path
. join
( train_dir
, 'smile' )
os
. mkdir
( train_smile_dir
)
train_unsmile_dir
= os
. path
. join
( train_dir
, 'unsmile' )
validation_smile_dir
= os
. path
. join
( validation_dir
, 'smile' )
os
. mkdir
( validation_smile_dir
)
validation_unsmile_dir
= os
. path
. join
( validation_dir
, 'unsmile' )
os
. mkdir
( validation_unsmile_dir
)
test_smile_dir
= os
. path
. join
( test_dir
, 'smile' )
os
. mkdir
( test_smile_dir
)
test_unsmile_dir
= os
. path
. join
( test_dir
, 'unsmile' )
os
. mkdir
( test_unsmile_dir
)
分配數(shù)據(jù)集，可以使用人為劃分和代碼劃分 進(jìn)行一次檢查，計(jì)算每個(gè)分組中有多少張照片(訓(xùn)練/驗(yàn)證/測試) print ( 'total training smile images:' , len ( os
. listdir
( train_smile_dir
) ) )
total training smile images: 1000
print ( 'total testing smile images:' , len ( os
. listdir
( test_smile_dir
) ) )
total testing smile images: 300
print ( 'total training unsmile images:' , len ( os
. listdir
( train_unsmile_dir
) ) )
total training unsmile images: 1000
print ( 'total validation smile images:' , len ( os
. listdir
( validation_smile_dir
) ) )
total validation smile images: 300
print ( 'total testing unsmile images:' , len ( os
. listdir
( test_unsmile_dir
) ) )
total testing unsmile images: 300
print ( 'total validation unsmile images:' , len ( os
. listdir
( validation_unsmile_dir
) ) )
total validation unsmile images: 300
有2000個(gè)訓(xùn)練圖像，然后是600個(gè)驗(yàn)證圖像，600個(gè)測試圖像，其中每個(gè)分類都有相同數(shù)量的樣本，是一個(gè)平衡的二元分類問題，意味著分類準(zhǔn)確度將是合適的度量標(biāo)準(zhǔn)。 3.網(wǎng)絡(luò)模型
卷積網(wǎng)絡(luò)(convnets)將是一組交替的Conv2D(具有relu激活)和MaxPooling2D層。從大小150x150（有點(diǎn)任意選擇）的輸入開始，我們最終得到了尺寸為７x7的Flatten層之前的特征圖。 注意特征圖的深度在網(wǎng)絡(luò)中逐漸增加(從32到128)，而特征圖的大小正在減少(從148x148到7x7)。這是一個(gè)你將在幾乎所有的卷積網(wǎng)絡(luò)(convnets)結(jié)構(gòu)中會看到的模式。 由于我們正在處理二元分類問題，所以我們用一個(gè)神經(jīng)元（一個(gè)大小為１的密集層(Dense))和一個(gè)sigmoid激活函數(shù)來結(jié)束網(wǎng)絡(luò)。該神經(jīng)元將會被用來查看圖像歸屬于那一類或另一類的概率。 from keras
import layers
from keras
import models
model
= models
. Sequential
( )
model
. add
( layers
. Conv2D
( 32 , ( 3 , 3 ) , activation
= 'relu' , input_shape
= ( 150 , 150 , 3 ) ) )
model
. add
( layers
. MaxPooling2D
( ( 2 , 2 ) ) )
model
. add
( layers
. Conv2D
( 64 , ( 3 , 3 ) , activation
= 'relu' ) )
model
. add
( layers
. MaxPooling2D
( ( 2 , 2 ) ) )
model
. add
( layers
. Conv2D
( 128 , ( 3 , 3 ) , activation
= 'relu' ) )
model
. add
( layers
. MaxPooling2D
( ( 2 , 2 ) ) )
model
. add
( layers
. Conv2D
( 128 , ( 3 , 3 ) , activation
= 'relu' ) )
model
. add
( layers
. MaxPooling2D
( ( 2 , 2 ) ) )
model
. add
( layers
. Flatten
( ) )
model
. add
( layers
. Dense
( 512 , activation
= 'relu' ) )
model
. add
( layers
. Dense
( 1 , activation
= 'sigmoid' ) )
看特征圖的尺寸如何隨著每個(gè)連續(xù)的圖層而改變，打印網(wǎng)絡(luò)結(jié)構(gòu)
model
. summary
( )
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
conv2d_5 (Conv2D) (None, 148, 148, 32) 896
_________________________________________________________________
max_pooling2d_5 (MaxPooling2 (None, 74, 74, 32) 0
_________________________________________________________________
conv2d_6 (Conv2D) (None, 72, 72, 64) 18496
_________________________________________________________________
max_pooling2d_6 (MaxPooling2 (None, 36, 36, 64) 0
_________________________________________________________________
conv2d_7 (Conv2D) (None, 34, 34, 128) 73856
_________________________________________________________________
max_pooling2d_7 (MaxPooling2 (None, 17, 17, 128) 0
_________________________________________________________________
conv2d_8 (Conv2D) (None, 15, 15, 128) 147584
_________________________________________________________________
max_pooling2d_8 (MaxPooling2 (None, 7, 7, 128) 0
_________________________________________________________________
flatten_2 (Flatten) (None, 6272) 0
_________________________________________________________________
dense_3 (Dense) (None, 512) 3211776
_________________________________________________________________
dense_4 (Dense) (None, 1) 513
=================================================================
Total params: 3,453,121
Trainable params: 3,453,121
Non-trainable params: 0
_________________________________________________________________
在編譯步驟里，使用RMSprop優(yōu)化器。由于用一個(gè)單一的神經(jīng)元(Sigmoid的激活函數(shù)）結(jié)束了網(wǎng)絡(luò)，將使用二進(jìn)制交叉熵(binary crossentropy)作為損失函數(shù) from keras
import optimizersmodel
. compile ( loss
= 'binary_crossentropy' , optimizer
= optimizers
. RMSprop
( lr
= 1e - 4 ) , metrics
= [ 'acc' ] )
4.資料預(yù)處理
網(wǎng)絡(luò)的預(yù)處理步驟： 讀入圖像 將JPEG內(nèi)容解碼為RGB網(wǎng)格的像素 將像素值(0和255之間)重新縮放到[0,1]間隔 數(shù)據(jù)應(yīng)該先被格式化成適當(dāng)?shù)念A(yù)處理浮點(diǎn)張量，然后才能輸入到神經(jīng)網(wǎng)絡(luò)中 from keras
. preprocessing
. image
import ImageDataGenerator
train_datagen
= ImageDataGenerator
( rescale
= 1 . / 255 )
validation_datagen
= ImageDataGenerator
( rescale
= 1 . / 255 )
test_datagen
= ImageDataGenerator
( rescale
= 1 . / 255 )
train_generator
= train_datagen
. flow_from_directory
( train_dir
, target_size
= ( 150 , 150 ) , batch_size
= 20 , class_mode
= 'binary' )
validation_generator
= test_datagen
. flow_from_directory
( validation_dir
, target_size
= ( 150 , 150 ) , batch_size
= 20 , class_mode
= 'binary' )
test_generator
= test_datagen
. flow_from_directory
( test_dir
, target_size
= ( 150 , 150 ) , batch_size
= 20 , class_mode
= 'binary' )
Found 2000 images belonging to 2 classes.
Found 600 images belonging to 2 classes.
Found 600 images belonging to 2 classes.
圖像張量生成器(generator)的輸出，它產(chǎn)生150x150 RGB圖像(形狀"(20,150,150,3)")和二進(jìn)制標(biāo)簽(形狀"(20,)")的批次張量。20是每個(gè)批次中的樣品數(shù)(批次大小) for data_batch
, labels_batch
in train_generator
: print ( 'data batch shape:' , data_batch
. shape
) print ( 'labels batch shape:' , labels_batch
) break
data batch shape: (20, 150, 150, 3)
labels batch shape: [ 1. 1. 0. 0. 1. 0. 0. 0. 1. 0. 0. 0. 0. 0. 1. 0. 0. 0.0. 0.]
train_generator
. class_indices
{'smile': 0, 'unsmile': 1}
5.開始訓(xùn)練
其中epochs值越大，花費(fèi)的時(shí)間就越久、訓(xùn)練的精度更高，我電腦性能不好，運(yùn)行了很久… …
history
= model
. fit_generator
( train_generator
, steps_per_epoch
= 100 , epochs
= 10 , validation_data
= validation_generator
, validation_steps
= 50 )
Epoch 1/10
100/100 [==============================] - 227s 2s/step - loss: 0.6776 - acc: 0.5740 - val_loss: 0.6745 - val_acc: 0.5660
Epoch 2/10
100/100 [==============================] - 230s 2s/step - loss: 0.6422 - acc: 0.6520 - val_loss: 0.7091 - val_acc: 0.5290
Epoch 3/10
100/100 [==============================] - 222s 2s/step - loss: 0.5889 - acc: 0.7020 - val_loss: 0.5711 - val_acc: 0.7530
Epoch 4/10
100/100 [==============================] - 192s 2s/step - loss: 0.5251 - acc: 0.7575 - val_loss: 0.5592 - val_acc: 0.7330
Epoch 5/10
100/100 [==============================] - 191s 2s/step - loss: 0.4854 - acc: 0.7825 - val_loss: 0.5250 - val_acc: 0.7550
Epoch 6/10
100/100 [==============================] - 184s 2s/step - loss: 0.4503 - acc: 0.8015 - val_loss: 0.5111 - val_acc: 0.7980
Epoch 7/10
100/100 [==============================] - 183s 2s/step - loss: 0.4111 - acc: 0.8255 - val_loss: 0.5376 - val_acc: 0.7500
Epoch 8/10
100/100 [==============================] - 189s 2s/step - loss: 0.3748 - acc: 0.8380 - val_loss: 0.4906 - val_acc: 0.7850
Epoch 9/10
100/100 [==============================] - 188s 2s/step - loss: 0.3493 - acc: 0.8590 - val_loss: 0.4397 - val_acc: 0.8170
Epoch 10/10
100/100 [==============================] - 186s 2s/step - loss: 0.3177 - acc: 0.8670 - val_loss: 0.4265 - val_acc: 0.8400
model
. save
( 'mangoout/smileAndUnsmile_1.h5' )
使用圖表來繪制在訓(xùn)練過程中模型對訓(xùn)練和驗(yàn)證數(shù)據(jù)的損失(loss)和準(zhǔn)確性(accuracy)數(shù)據(jù) import matplotlib
. pyplot
as plt
acc
= history
. history
[ 'acc' ]
val_acc
= history
. history
[ 'val_acc' ]
loss
= history
. history
[ 'loss' ]
val_loss
= history
. history
[ 'val_loss' ]
epochs
= range ( len ( acc
) )
plt
. plot
( epochs
, acc
, 'bo' , label
= 'Training acc' )
plt
. plot
( epochs
, val_acc
, 'b' , label
= 'Validation acc' )
plt
. title
( 'Training and validation accuracy' )
plt
. legend
( )
plt
. figure
( )
plt
. plot
( epochs
, loss
, 'bo' , label
= 'Training loss' )
plt
. plot
( epochs
, val_loss
, 'b' , label
= 'Validation loss' )
plt
. title
( 'Training and validation loss' )
plt
. legend
( )
plt
. show
( )
可能由于我們提前對圖片進(jìn)行了裁剪篩選操作，過度擬合這里不是太能體現(xiàn)出來，訓(xùn)練精確度隨著時(shí)間線性增長，驗(yàn)證精度也在增長。 驗(yàn)證損失和訓(xùn)練損失都在線性上保持下降直到接近０。 💥 為了進(jìn)一步優(yōu)化，下面會引入一個(gè)新的、特定于電腦視覺影像，并在使用深度學(xué)習(xí)模型處理圖像時(shí)幾乎普遍使用的技巧:數(shù)據(jù)擴(kuò)充(data augmentation)
6.使用數(shù)據(jù)填充
數(shù)據(jù)增加采用從現(xiàn)有訓(xùn)練樣本生成更多訓(xùn)練數(shù)據(jù)的方法，通過產(chǎn)生可信的圖像的多個(gè)隨機(jī)變換來"增加"樣本。目標(biāo)是在訓(xùn)練的時(shí)候，我們的模型永遠(yuǎn)不會再看到完全相同的畫面兩次。這有助于模型學(xué)習(xí)到數(shù)據(jù)的更多方面，并更好地推廣。 在keras中，可以通過配置對ImageDataGenerator實(shí)例讀取的圖像執(zhí)行多個(gè)隨機(jī)變換來完成
datagen
= ImageDataGenerator
( rotation_range
= 40 , width_shift_range
= 0.2 , height_shift_range
= 0.2 , shear_range
= 0.2 , zoom_range
= 0.2 , horizontal_flip
= True , fill_mode
= 'nearest' )
這里列出一些可用的選項(xiàng)(更多選項(xiàng)，可以參考keras文檔），快速看一下這些參數(shù): rotation_range 是以度（0-180)為單位的值，它是隨機(jī)旋轉(zhuǎn)圖片的范圍 width_shift 和 height_shift 是范圍(占總寬度或高度的一小部分),用于縱向或橫向隨機(jī)轉(zhuǎn)換圖片 shear_range 用于隨機(jī)剪切變換 zoom_range 用于隨機(jī)放大圖片內(nèi)容 horizontal_flip 用于在沒有水平不對稱假設(shè)（例如真實(shí)世界圖片）的情況下水平地隨機(jī)翻轉(zhuǎn)一半圖像 fill_mode 是用于填充新創(chuàng)建的像素的策略，可以在旋轉(zhuǎn)或?qū)?高移位后顯示 import matplotlib
. pyplot
as plt
from keras
. preprocessing
import image
fnames
= [ os
. path
. join
( train_smile_dir
, fname
) for fname
in os
. listdir
( train_smile_dir
) ]
img_path
= fnames
[ 3 ]
img
= image
. load_img
( img_path
, target_size
= ( 150 , 150 ) )
x
= image
. img_to_array
( img
)
x
= x
. reshape
( ( 1 , ) + x
. shape
)
i
= 0
for batch
in datagen
. flow
( x
, batch_size
= 1 ) : plt
. figure
( i
) imgplot
= plt
. imshow
( image
. array_to_img
( batch
[ 0 ] ) ) i
+= 1 if i
% 4 == 0 : break
plt
. show
( )
如果我們使用這種數(shù)據(jù)增強(qiáng)配置來訓(xùn)練一個(gè)新的網(wǎng)絡(luò)，我們的網(wǎng)絡(luò)將永遠(yuǎn)不會看到相同重復(fù)的輸入。 然而，它看到的輸入仍然是相互關(guān)聯(lián)的，因?yàn)樗鼈儊碜陨倭康脑紙D像 - 我們不能產(chǎn)生新的信息，我們只能重新混合現(xiàn)有的信息。 因此，這可能不足以完全擺脫過度擬合(overfitting)。為了進(jìn)一步克服過度擬合(overfitting)，我們還將在密集連接(densely-connected)的分類器之前添加一個(gè)Dropout層。
model
= models
. Sequential
( )
model
. add
( layers
. Conv2D
( 32 , ( 3 , 3 ) , activation
= 'relu' , input_shape
= ( 150 , 150 , 3 ) ) )
model
. add
( layers
. MaxPooling2D
( ( 2 , 2 ) ) )
model
. add
( layers
. Conv2D
( 64 , ( 3 , 3 ) , activation
= 'relu' ) )
model
. add
( layers
. MaxPooling2D
( ( 2 , 2 ) ) )
model
. add
( layers
. Conv2D
( 128 , ( 3 , 3 ) , activation
= 'relu' ) )
model
. add
( layers
. MaxPooling2D
( ( 2 , 2 ) ) )
model
. add
( layers
. Conv2D
( 128 , ( 3 , 3 ) , activation
= 'relu' ) )
model
. add
( layers
. MaxPooling2D
( ( 2 , 2 ) ) )
model
. add
( layers
. Flatten
( ) )
model
. add
( layers
. Dropout
( 0.5 ) )
model
. add
( layers
. Dense
( 512 , activation
= 'relu' ) )
model
. add
( layers
. Dense
( 1 , activation
= 'sigmoid' ) )
model
. compile ( loss
= 'binary_crossentropy' , optimizer
= optimizers
. RMSprop
( lr
= 1e - 4 ) , metrics
= [ 'acc' ] )
使用數(shù)據(jù)填充(data augmentation)和dropout來訓(xùn)練我們的網(wǎng)絡(luò)
train_datagen
= ImageDataGenerator
( rescale
= 1 . / 255 , rotation_range
= 40 , width_shift_range
= 0.2 , height_shift_range
= 0.2 , shear_range
= 0.2 , zoom_range
= 0.2 , horizontal_flip
= True , )
test_datagen
= ImageDataGenerator
( rescale
= 1 . / 255 ) train_generator
= train_datagen
. flow_from_directory
( train_dir
, target_size
= ( 150 , 150 ) , batch_size
= 32 , class_mode
= 'binary' ) validation_generator
= test_datagen
. flow_from_directory
( validation_dir
, target_size
= ( 150 , 150 ) , batch_size
= 32 , class_mode
= 'binary' ) history
= model
. fit_generator
( train_generator
, steps_per_epoch
= 100 , epochs
= 60 , validation_data
= validation_generator
, validation_steps
= 50 )
Found 2000 images belonging to 2 classes.
Found 600 images belonging to 2 classes.
Epoch 1/60
100/100 [==============================] - 377s 4s/step - loss: 0.6905 - acc: 0.5312 - val_loss: 0.6815 - val_acc: 0.5587
Epoch 2/60
100/100 [==============================] - 303s 3s/step - loss: 0.6849 - acc: 0.5550 - val_loss: 0.6768 - val_acc: 0.5852
Epoch 3/60
100/100 [==============================] - 314s 3s/step - loss: 0.6790 - acc: 0.5866 - val_loss: 0.6748 - val_acc: 0.5777
Epoch 4/60
100/100 [==============================] - 312s 3s/step - loss: 0.6774 - acc: 0.5806 - val_loss: 0.6722 - val_acc: 0.5821
Epoch 5/60
100/100 [==============================] - 303s 3s/step - loss: 0.6696 - acc: 0.5869 - val_loss: 0.6734 - val_acc: 0.5909
Epoch 6/60
100/100 [==============================] - 300s 3s/step - loss: 0.6665 - acc: 0.5956 - val_loss: 0.6679 - val_acc: 0.6029
Epoch 7/60
100/100 [==============================] - 309s 3s/step - loss: 0.6719 - acc: 0.5853 - val_loss: 0.6836 - val_acc: 0.5814
Epoch 8/60
100/100 [==============================] - 299s 3s/step - loss: 0.6628 - acc: 0.6031 - val_loss: 0.7166 - val_acc: 0.5600
Epoch 9/60
100/100 [==============================] - 310s 3s/step - loss: 0.6676 - acc: 0.6009 - val_loss: 0.6723 - val_acc: 0.6048
Epoch 10/60
100/100 [==============================] - 309s 3s/step - loss: 0.6629 - acc: 0.6100 - val_loss: 0.6546 - val_acc: 0.6288
Epoch 11/60
100/100 [==============================] - 300s 3s/step - loss: 0.6558 - acc: 0.6163 - val_loss: 0.6744 - val_acc: 0.6073
Epoch 12/60
100/100 [==============================] - 300s 3s/step - loss: 0.6575 - acc: 0.6216 - val_loss: 0.6425 - val_acc: 0.6572
Epoch 13/60
100/100 [==============================] - 307s 3s/step - loss: 0.6541 - acc: 0.6294 - val_loss: 0.7095 - val_acc: 0.5960
Epoch 14/60
100/100 [==============================] - 303s 3s/step - loss: 0.6429 - acc: 0.6400 - val_loss: 0.6381 - val_acc: 0.6414
Epoch 15/60
100/100 [==============================] - 310s 3s/step - loss: 0.6427 - acc: 0.6300 - val_loss: 0.6297 - val_acc: 0.6723
Epoch 16/60
100/100 [==============================] - 308s 3s/step - loss: 0.6404 - acc: 0.6369 - val_loss: 0.6254 - val_acc: 0.6667
Epoch 17/60
100/100 [==============================] - 301s 3s/step - loss: 0.6367 - acc: 0.6500 - val_loss: 0.6145 - val_acc: 0.6408
Epoch 18/60
100/100 [==============================] - 301s 3s/step - loss: 0.6246 - acc: 0.6450 - val_loss: 0.5991 - val_acc: 0.6831
Epoch 19/60
100/100 [==============================] - 307s 3s/step - loss: 0.6230 - acc: 0.6625 - val_loss: 0.5956 - val_acc: 0.7052
Epoch 20/60
100/100 [==============================] - 303s 3s/step - loss: 0.6062 - acc: 0.6725 - val_loss: 0.5812 - val_acc: 0.6951
Epoch 21/60
100/100 [==============================] - 301s 3s/step - loss: 0.6094 - acc: 0.6647 - val_loss: 0.5640 - val_acc: 0.7033
Epoch 22/60
100/100 [==============================] - 303s 3s/step - loss: 0.6089 - acc: 0.6794 - val_loss: 0.5698 - val_acc: 0.6774
Epoch 23/60
100/100 [==============================] - 305s 3s/step - loss: 0.5846 - acc: 0.6869 - val_loss: 0.5458 - val_acc: 0.7311
Epoch 24/60
100/100 [==============================] - 300s 3s/step - loss: 0.5782 - acc: 0.7022 - val_loss: 0.5380 - val_acc: 0.7551
Epoch 25/60
100/100 [==============================] - 303s 3s/step - loss: 0.5676 - acc: 0.7112 - val_loss: 0.5731 - val_acc: 0.7109
Epoch 26/60
100/100 [==============================] - 305s 3s/step - loss: 0.5711 - acc: 0.7134 - val_loss: 0.5303 - val_acc: 0.7121
Epoch 27/60
100/100 [==============================] - 315s 3s/step - loss: 0.5565 - acc: 0.7166 - val_loss: 0.4785 - val_acc: 0.7942
Epoch 28/60
100/100 [==============================] - 307s 3s/step - loss: 0.5462 - acc: 0.7269 - val_loss: 0.5904 - val_acc: 0.7052
Epoch 29/60
100/100 [==============================] - 309s 3s/step - loss: 0.5367 - acc: 0.7309 - val_loss: 0.4601 - val_acc: 0.7847
Epoch 30/60
100/100 [==============================] - 304s 3s/step - loss: 0.5283 - acc: 0.7359 - val_loss: 0.5043 - val_acc: 0.7532
Epoch 31/60
100/100 [==============================] - 302s 3s/step - loss: 0.5202 - acc: 0.7472 - val_loss: 0.5292 - val_acc: 0.7424
Epoch 32/60
100/100 [==============================] - 308s 3s/step - loss: 0.5147 - acc: 0.7531 - val_loss: 0.5043 - val_acc: 0.7860
Epoch 33/60
100/100 [==============================] - 305s 3s/step - loss: 0.5100 - acc: 0.7434 - val_loss: 0.4506 - val_acc: 0.7955
Epoch 34/60
100/100 [==============================] - 303s 3s/step - loss: 0.5067 - acc: 0.7628 - val_loss: 0.4423 - val_acc: 0.7803
Epoch 35/60
100/100 [==============================] - 303s 3s/step - loss: 0.4937 - acc: 0.7591 - val_loss: 0.4281 - val_acc: 0.8037
Epoch 36/60
100/100 [==============================] - 300s 3s/step - loss: 0.4903 - acc: 0.7619 - val_loss: 0.4191 - val_acc: 0.8125
Epoch 37/60
100/100 [==============================] - 299s 3s/step - loss: 0.4704 - acc: 0.7769 - val_loss: 0.4266 - val_acc: 0.8213
Epoch 38/60
100/100 [==============================] - 307s 3s/step - loss: 0.4811 - acc: 0.7778 - val_loss: 0.4196 - val_acc: 0.8239
Epoch 39/60
100/100 [==============================] - 312s 3s/step - loss: 0.4722 - acc: 0.7753 - val_loss: 0.4366 - val_acc: 0.7992
Epoch 40/60
100/100 [==============================] - 301s 3s/step - loss: 0.4694 - acc: 0.7797 - val_loss: 0.4597 - val_acc: 0.7879
Epoch 41/60
100/100 [==============================] - 304s 3s/step - loss: 0.4658 - acc: 0.7866 - val_loss: 0.4021 - val_acc: 0.8239
Epoch 42/60
100/100 [==============================] - 303s 3s/step - loss: 0.4700 - acc: 0.7859 - val_loss: 0.4271 - val_acc: 0.8100
Epoch 43/60
100/100 [==============================] - 300s 3s/step - loss: 0.4591 - acc: 0.7850 - val_loss: 0.4687 - val_acc: 0.7898
Epoch 44/60
100/100 [==============================] - 308s 3s/step - loss: 0.4592 - acc: 0.7847 - val_loss: 0.4136 - val_acc: 0.8169
Epoch 45/60
100/100 [==============================] - 311s 3s/step - loss: 0.4449 - acc: 0.8031 - val_loss: 0.4427 - val_acc: 0.7784
Epoch 46/60
100/100 [==============================] - 299s 3s/step - loss: 0.4505 - acc: 0.7897 - val_loss: 0.4030 - val_acc: 0.8220
Epoch 47/60
100/100 [==============================] - 306s 3s/step - loss: 0.4515 - acc: 0.7978 - val_loss: 0.4324 - val_acc: 0.7948
Epoch 48/60
100/100 [==============================] - 303s 3s/step - loss: 0.4485 - acc: 0.8025 - val_loss: 0.4979 - val_acc: 0.7544
Epoch 49/60
100/100 [==============================] - 303s 3s/step - loss: 0.4310 - acc: 0.8137 - val_loss: 0.4203 - val_acc: 0.8100
Epoch 50/60
100/100 [==============================] - 309s 3s/step - loss: 0.4327 - acc: 0.8044 - val_loss: 0.4064 - val_acc: 0.8194
Epoch 51/60
100/100 [==============================] - 313s 3s/step - loss: 0.4308 - acc: 0.8091 - val_loss: 0.4266 - val_acc: 0.7891
Epoch 52/60
100/100 [==============================] - 307s 3s/step - loss: 0.4321 - acc: 0.7997 - val_loss: 0.4612 - val_acc: 0.8169
Epoch 53/60
100/100 [==============================] - 303s 3s/step - loss: 0.4316 - acc: 0.8081 - val_loss: 0.4235 - val_acc: 0.7942
Epoch 54/60
100/100 [==============================] - 302s 3s/step - loss: 0.4226 - acc: 0.8135 - val_loss: 0.4083 - val_acc: 0.8302
Epoch 55/60
100/100 [==============================] - 300s 3s/step - loss: 0.4240 - acc: 0.8062 - val_loss: 0.3863 - val_acc: 0.8340
Epoch 56/60
100/100 [==============================] - 308s 3s/step - loss: 0.4139 - acc: 0.8184 - val_loss: 0.4100 - val_acc: 0.8232
Epoch 57/60
100/100 [==============================] - 303s 3s/step - loss: 0.4149 - acc: 0.8159 - val_loss: 0.3911 - val_acc: 0.8239
Epoch 58/60
100/100 [==============================] - 305s 3s/step - loss: 0.4171 - acc: 0.8113 - val_loss: 0.3851 - val_acc: 0.8321
Epoch 59/60
100/100 [==============================] - 304s 3s/step - loss: 0.4221 - acc: 0.8141 - val_loss: 0.4243 - val_acc: 0.7891
Epoch 60/60
100/100 [==============================] - 303s 3s/step - loss: 0.4019 - acc: 0.8262 - val_loss: 0.3919 - val_acc: 0.8371
train_generator
. class_indices
{'smile': 0, 'unsmile': 1}
model
. save
( 'mangoout/smileAndUnsmile_2.h5' )
保存模型文件夾顯示效果 繪制數(shù)據(jù)增強(qiáng)后的訓(xùn)練集與驗(yàn)證集的精確度與損失度的圖形，看一遍結(jié)果
acc
= history
. history
[ 'acc' ]
val_acc
= history
. history
[ 'val_acc' ]
loss
= history
. history
[ 'loss' ]
val_loss
= history
. history
[ 'val_loss' ] epochs
= range ( len ( acc
) ) plt
. plot
( epochs
, acc
, 'bo' , label
= 'Training acc' )
plt
. plot
( epochs
, val_acc
, 'b' , label
= 'Validation acc' )
plt
. title
( 'Training and validation accuracy' )
plt
. legend
( ) plt
. figure
( ) plt
. plot
( epochs
, loss
, 'bo' , label
= 'Training loss' )
plt
. plot
( epochs
, val_loss
, 'b' , label
= 'Validation loss' )
plt
. title
( 'Training and validation loss' )
plt
. legend
( ) plt
. show
( )
由于數(shù)據(jù)增加(data augmentation)和dropout使用，不再有過度擬合(overfitting)的問題；訓(xùn)練曲線相當(dāng)密切地跟隨著驗(yàn)證曲線。訓(xùn)練精度和驗(yàn)證精度經(jīng)過60個(gè)循環(huán)接近85%。 通過進(jìn)一步利用正規(guī)化技術(shù)，及調(diào)整網(wǎng)絡(luò)參數(shù)(例如每個(gè)卷積層的濾波器數(shù)量或網(wǎng)絡(luò)層數(shù))，可以獲得更好的準(zhǔn)確度。 7.對人臉微笑與否的模型訓(xùn)練
判斷的第一張圖片（D:/mango/mangotest.jpg路徑下）
from keras
. preprocessing
import image
from keras
. models
import load_model
import numpy
as np
model
= load_model
( 'mangoout/smileAndUnsmile_2.h5' )
img_path
= 'D:/mango/mangotest.jpg'
img
= image
. load_img
( img_path
, target_size
= ( 150 , 150 ) ) img_tensor
= image
. img_to_array
( img
) / 255.0
img_tensor
= np
. expand_dims
( img_tensor
, axis
= 0 )
prediction
= model
. predict
( img_tensor
)
print ( prediction
)
if prediction
[ 0 ] [ 0 ] > 0.5 : result
= '非笑臉'
else : result
= '笑臉'
print ( result
)
[[ 0.00120554]]
笑臉
判斷的第二張圖片（D:/mango/mangotest2.jpg路徑下） 這是我假期無聊時(shí)用美圖拍的，哈哈… … ?? from keras
. preprocessing
import image
from keras
. models
import load_model
import numpy
as np
model
= load_model
( 'mangoout/smileAndUnsmile_2.h5' )
img_path
= 'D:/mango/mangotest2.jpg'
img
= image
. load_img
( img_path
, target_size
= ( 150 , 150 ) ) img_tensor
= image
. img_to_array
( img
) / 255.0
img_tensor
= np
. expand_dims
( img_tensor
, axis
= 0 )
prediction
= model
. predict
( img_tensor
)
print ( prediction
)
if prediction
[ 0 ] [ 0 ] > 0.5 : result
= '非笑臉'
else : result
= '笑臉'
print ( result
)
[[ 0.69130015]]
非笑臉
判斷的第三張圖片（D:/mango/mengmengsmile.jpg路徑下） 她自己拍的，順便被我拿來做作業(yè)，哈哈… … from keras
. preprocessing
import image
from keras
. models
import load_model
import numpy
as np
model
= load_model
( 'mangoout/smileAndUnsmile_1.h5' )
img_path
= 'D:/mango/mengmengsmile.jpg'
img
= image
. load_img
( img_path
, target_size
= ( 150 , 150 ) ) img_tensor
= image
. img_to_array
( img
) / 255.0
img_tensor
= np
. expand_dims
( img_tensor
, axis
= 0 )
prediction
= model
. predict
( img_tensor
)
print ( prediction
)
if prediction
[ 0 ] [ 0 ] > 0.5 : result
= '非笑臉'
else : result
= '笑臉'
print ( result
)
[[ 0.03113164]]
笑臉
可以看見，判斷為笑臉是正確的（但是有一些圖片中人臉部分在整張圖片的占比不是在特別多的話，它判斷出來的準(zhǔn)確度就比較小，甚至出現(xiàn)判斷錯(cuò)誤的情況，但是一般大頭照判斷出來的準(zhǔn)確度還是很高的。） 三、卷積神經(jīng)網(wǎng)絡(luò)（CNN）對口罩?jǐn)?shù)據(jù)集正負(fù)樣本的劃分、模型訓(xùn)練和測試的過程，輸出模型訓(xùn)練精度和測試精度，完成對口罩佩戴與否的模型訓(xùn)練
1.數(shù)據(jù)集準(zhǔn)備
下載口罩?jǐn)?shù)據(jù)集，把它解壓到相應(yīng)的目錄（我放在了D:\mangomask目錄下） 解壓后原始有口罩的數(shù)據(jù)集顯示效果 解壓后原始沒有口罩的數(shù)據(jù)集顯示效果 將正樣本（有口罩）數(shù)據(jù)集重命名為連續(xù)序列，以便后面調(diào)整 import os
path
= "D:/mangomask/mask/have_mask"
filelist
= os
. listdir
( path
)
count
= 1000
for file in filelist
: Olddir
= os
. path
. join
( path
, file ) if os
. path
. isdir
( Olddir
) : continue filename
= os
. path
. splitext
( file ) [ 0 ] filetype
= os
. path
. splitext
( file ) [ 1 ] Newdir
= os
. path
. join
( path
, str ( count
) + filetype
) os
. rename
( Olddir
, Newdir
) count
+= 1
對數(shù)據(jù)集重命名后，人臉正樣本（有口罩）數(shù)據(jù)集顯示效果如下 將負(fù)樣本（沒有口罩）數(shù)據(jù)集重命名為連續(xù)序列，以便后面調(diào)整 import os
path
= "D:/mangomask/mask/no_mask"
filelist
= os
. listdir
( path
)
count
= 10000
for file in filelist
: Olddir
= os
. path
. join
( path
, file ) if os
. path
. isdir
( Olddir
) : continue filename
= os
. path
. splitext
( file ) [ 0 ] filetype
= os
. path
. splitext
( file ) [ 1 ] Newdir
= os
. path
. join
( path
, str ( count
) + filetype
) os
. rename
( Olddir
, Newdir
) count
+= 1
對數(shù)據(jù)集重命名后，人臉負(fù)樣本（沒有口罩）數(shù)據(jù)集顯示效果如下 正負(fù)樣本數(shù)據(jù)集像素處理 正樣本（有口罩）數(shù)據(jù)集的像素設(shè)置為 20x20，模型訓(xùn)練精度更高；
1.調(diào)整正樣本（有口罩）像素
import pandas
as pd
import cv2
for n
in range ( 1000 , 1606 ) : path
= 'D:/mangomask/mask/have_mask/' + str ( n
) + '.jpg' img
= cv2
. imread
( path
) img
= cv2
. resize
( img
, ( 20 , 20 ) ) cv2
. imwrite
( 'D:/mangomask/mask/have_mask/' + str ( n
) + '.jpg' , img
) n
+= 1
調(diào)整正樣本（有口罩）像素后，數(shù)據(jù)集圖像顯示效果 import pandas
as pd
import cv2
for n
in range ( 10000 , 11790 ) : path
= 'D:/mangomask/mask/no_mask/' + str ( n
) + '.jpg' img
= cv2
. imread
( path
) img
= cv2
. resize
( img
, ( 80 , 80 ) ) cv2
. imwrite
( 'D:/mangomask/mask/no_mask/' + str ( n
) + '.jpg' , img
) n
+= 1
調(diào)整負(fù)樣本（沒有口罩）像素后，數(shù)據(jù)集圖像顯示效果 劃分?jǐn)?shù)據(jù)集（在當(dāng)前寫代碼的同級目錄下會產(chǎn)生一個(gè)maskout的文件夾，包括 train :訓(xùn)練集；validation：驗(yàn)證集；test：測試集。）
original_dataset_dir
= 'D://mangomask//mask' base_dir
= 'maskout'
os
. mkdir
( base_dir
) train_dir
= os
. path
. join
( base_dir
, 'train' )
os
. mkdir
( train_dir
)
validation_dir
= os
. path
. join
( base_dir
, 'validation' )
os
. mkdir
( validation_dir
)
test_dir
= os
. path
. join
( base_dir
, 'test' )
os
. mkdir
( test_dir
) train_havemask_dir
= os
. path
. join
( train_dir
, 'have_mask' )
os
. mkdir
( train_havemask_dir
) train_nomask_dir
= os
. path
. join
( train_dir
, 'no_mask' )
os
. mkdir
( train_nomask_dir
) validation_havemask_dir
= os
. path
. join
( validation_dir
, 'have_mask' )
os
. mkdir
( validation_havemask_dir
) validation_nomask_dir
= os
. path
. join
( validation_dir
, 'no_mask' )
os
. mkdir
( validation_nomask_dir
) test_havemask_dir
= os
. path
. join
( test_dir
, 'have_mask' )
os
. mkdir
( test_havemask_dir
) test_nomask_dir
= os
. path
. join
( test_dir
, 'no_mask' )
os
. mkdir
( test_nomask_dir
)
分配數(shù)據(jù)集，可以使用人為劃分和代碼劃分 進(jìn)行一次檢查，計(jì)算每個(gè)分組中有多少張照片(訓(xùn)練/驗(yàn)證/測試) print ( 'total training havemask images:' , len ( os
. listdir
( train_havemask_dir
) ) )
total training havemask images: 300
print ( 'total testing havemask images:' , len ( os
. listdir
( test_havemask_dir
) ) )
total testing havemask images: 150
print ( 'total training nomask images:' , len ( os
. listdir
( train_nomask_dir
) ) )
total training nomask images: 300
print ( 'total validation havemask images:' , len ( os
. listdir
( validation_havemask_dir
) ) )
total validation havemask images: 150
print ( 'total testing nomask images:' , len ( os
. listdir
( test_nomask_dir
) ) )
total testing nomask images: 150
print ( 'total validation nomask images:' , len ( os
. listdir
( validation_nomask_dir
) ) )
total validation nomask images: 150
有600個(gè)訓(xùn)練圖像，然后是300個(gè)驗(yàn)證圖像，300個(gè)測試圖像，其中每個(gè)分類都有相同數(shù)量的樣本，是一個(gè)平衡的二元分類問題，意味著分類準(zhǔn)確度將是合適的度量標(biāo)準(zhǔn)。 2.網(wǎng)絡(luò)模型
卷積網(wǎng)絡(luò)(convnets)將是一組交替的Conv2D(具有relu激活)和MaxPooling2D層。從大小150x150（有點(diǎn)任意選擇）的輸入開始，我們最終得到了尺寸為７x7的Flatten層之前的特征圖。 注意特征圖的深度在網(wǎng)絡(luò)中逐漸增加(從32到128)，而特征圖的大小正在減少(從148x148到7x7)。這是一個(gè)你將在幾乎所有的卷積網(wǎng)絡(luò)(convnets)結(jié)構(gòu)中會看到的模式。 由于我們正在處理二元分類問題，所以我們用一個(gè)神經(jīng)元（一個(gè)大小為１的密集層(Dense))和一個(gè)sigmoid激活函數(shù)來結(jié)束網(wǎng)絡(luò)。該神經(jīng)元將會被用來查看圖像歸屬于那一類或另一類的概率。 from keras
import layers
from keras
import modelsmodel
= models
. Sequential
( )
model
. add
( layers
. Conv2D
( 32 , ( 3 , 3 ) , activation
= 'relu' , input_shape
= ( 150 , 150 , 3 ) ) )
model
. add
( layers
. MaxPooling2D
( ( 2 , 2 ) ) )
model
. add
( layers
. Conv2D
( 64 , ( 3 , 3 ) , activation
= 'relu' ) )
model
. add
( layers
. MaxPooling2D
( ( 2 , 2 ) ) )
model
. add
( layers
. Conv2D
( 128 , ( 3 , 3 ) , activation
= 'relu' ) )
model
. add
( layers
. MaxPooling2D
( ( 2 , 2 ) ) )
model
. add
( layers
. Conv2D
( 128 , ( 3 , 3 ) , activation
= 'relu' ) )
model
. add
( layers
. MaxPooling2D
( ( 2 , 2 ) ) )
model
. add
( layers
. Flatten
( ) )
model
. add
( layers
. Dense
( 512 , activation
= 'relu' ) )
model
. add
( layers
. Dense
( 1 , activation
= 'sigmoid' ) )
看特征圖的尺寸如何隨著每個(gè)連續(xù)的圖層而改變，打印網(wǎng)絡(luò)結(jié)構(gòu)
model
. summary
( )
_________________________________________________________________
Layer (type) Output Shape Param #
=================================================================
conv2d_13 (Conv2D) (None, 148, 148, 32) 896
_________________________________________________________________
max_pooling2d_13 (MaxPooling (None, 74, 74, 32) 0
_________________________________________________________________
conv2d_14 (Conv2D) (None, 72, 72, 64) 18496
_________________________________________________________________
max_pooling2d_14 (MaxPooling (None, 36, 36, 64) 0
_________________________________________________________________
conv2d_15 (Conv2D) (None, 34, 34, 128) 73856
_________________________________________________________________
max_pooling2d_15 (MaxPooling (None, 17, 17, 128) 0
_________________________________________________________________
conv2d_16 (Conv2D) (None, 15, 15, 128) 147584
_________________________________________________________________
max_pooling2d_16 (MaxPooling (None, 7, 7, 128) 0
_________________________________________________________________
flatten_4 (Flatten) (None, 6272) 0
_________________________________________________________________
dense_7 (Dense) (None, 512) 3211776
_________________________________________________________________
dense_8 (Dense) (None, 1) 513
=================================================================
Total params: 3,453,121
Trainable params: 3,453,121
Non-trainable params: 0
_________________________________________________________________
在編譯步驟里，使用RMSprop優(yōu)化器。由于用一個(gè)單一的神經(jīng)元(Sigmoid的激活函數(shù)）結(jié)束了網(wǎng)絡(luò)，將使用二進(jìn)制交叉熵(binary crossentropy)作為損失函數(shù) from keras
import optimizersmodel
. compile ( loss
= 'binary_crossentropy' , optimizer
= optimizers
. RMSprop
( lr
= 1e - 4 ) , metrics
= [ 'acc' ] )
3.資料預(yù)處理
網(wǎng)絡(luò)的預(yù)處理步驟： 讀入圖像 將JPEG內(nèi)容解碼為RGB網(wǎng)格的像素 將像素值(0和255之間)重新縮放到[0,1]間隔 數(shù)據(jù)應(yīng)該先被格式化成適當(dāng)?shù)念A(yù)處理浮點(diǎn)張量，然后才能輸入到神經(jīng)網(wǎng)絡(luò)中 from keras
. preprocessing
. image
import ImageDataGenerator
train_datagen
= ImageDataGenerator
( rescale
= 1 . / 255 )
validation_datagen
= ImageDataGenerator
( rescale
= 1 . / 255 )
test_datagen
= ImageDataGenerator
( rescale
= 1 . / 255 )
train_generator
= train_datagen
. flow_from_directory
( train_dir
, target_size
= ( 150 , 150 ) , batch_size
= 20 , class_mode
= 'binary' )
validation_generator
= test_datagen
. flow_from_directory
( validation_dir
, target_size
= ( 150 , 150 ) , batch_size
= 20 , class_mode
= 'binary' )
test_generator
= test_datagen
. flow_from_directory
( test_dir
, target_size
= ( 150 , 150 ) , batch_size
= 20 , class_mode
= 'binary' )
Found 600 images belonging to 2 classes.
Found 300 images belonging to 2 classes.
Found 300 images belonging to 2 classes.
圖像張量生成器(generator)的輸出，它產(chǎn)生150x150 RGB圖像(形狀"(20,150,150,3)")和二進(jìn)制標(biāo)簽(形狀"(20,)")的批次張量。20是每個(gè)批次中的樣品數(shù)(批次大小) for data_batch
, labels_batch
in train_generator
: print ( 'data batch shape:' , data_batch
. shape
) print ( 'labels batch shape:' , labels_batch
) break
data batch shape: (20, 150, 150, 3)
labels batch shape: [ 1. 1. 1. 0. 1. 1. 0. 0. 0. 0. 1. 0. 0. 1. 1. 0. 1. 0.0. 0.]
4.開始訓(xùn)練
這里取epochs=10，其中epochs值越大，花費(fèi)的時(shí)間就越久、訓(xùn)練的精度更高，我電腦性能不好，運(yùn)行了很久… …
history
= model
. fit_generator
( train_generator
, steps_per_epoch
= 100 , epochs
= 10 , validation_data
= validation_generator
, validation_steps
= 50 )
Epoch 1/10
100/100 [==============================] - 218s 2s/step - loss: 0.2563 - acc: 0.8990 - val_loss: 0.1740 - val_acc: 0.9400
Epoch 2/10
100/100 [==============================] - 189s 2s/step - loss: 0.0862 - acc: 0.9700 - val_loss: 0.1294 - val_acc: 0.9540
Epoch 3/10
100/100 [==============================] - 190s 2s/step - loss: 0.0548 - acc: 0.9820 - val_loss: 0.1033 - val_acc: 0.9680
Epoch 4/10
100/100 [==============================] - 186s 2s/step - loss: 0.0325 - acc: 0.9880 - val_loss: 0.1132 - val_acc: 0.9620
Epoch 5/10
100/100 [==============================] - 192s 2s/step - loss: 0.0238 - acc: 0.9925 - val_loss: 0.0922 - val_acc: 0.9800
Epoch 6/10
100/100 [==============================] - 191s 2s/step - loss: 0.0132 - acc: 0.9965 - val_loss: 0.0950 - val_acc: 0.9710
Epoch 7/10
100/100 [==============================] - 189s 2s/step - loss: 0.0061 - acc: 0.9980 - val_loss: 0.1093 - val_acc: 0.9710
Epoch 8/10
100/100 [==============================] - 188s 2s/step - loss: 0.0025 - acc: 0.9995 - val_loss: 0.1305 - val_acc: 0.9690
Epoch 9/10
100/100 [==============================] - 185s 2s/step - loss: 0.0080 - acc: 0.9980 - val_loss: 0.1067 - val_acc: 0.9770
Epoch 10/10
100/100 [==============================] - 189s 2s/step - loss: 6.6883e-04 - acc: 1.0000 - val_loss: 0.1032 - val_acc: 0.9780
model
. save
( 'maskout/maskAndNomask_1.h5' )
使用圖表來繪制在訓(xùn)練過程中模型對訓(xùn)練和驗(yàn)證數(shù)據(jù)的損失(loss)和準(zhǔn)確性(accuracy)數(shù)據(jù) import matplotlib
. pyplot
as plt
acc
= history
. history
[ 'acc' ]
val_acc
= history
. history
[ 'val_acc' ]
loss
= history
. history
[ 'loss' ]
val_loss
= history
. history
[ 'val_loss' ] epochs
= range ( len ( acc
) )
plt
. plot
( epochs
, acc
, 'bo' , label
= 'Training acc' )
plt
. plot
( epochs
, val_acc
, 'b' , label
= 'Validation acc' )
plt
. title
( 'Training and validation accuracy' )
plt
. legend
( ) plt
. figure
( ) plt
. plot
( epochs
, loss
, 'bo' , label
= 'Training loss' )
plt
. plot
( epochs
, val_loss
, 'b' , label
= 'Validation loss' )
plt
. title
( 'Training and validation loss' )
plt
. legend
( ) plt
. show
( )
上面圖標(biāo)顯示了過度擬合(overfitting)的特征。我們的訓(xùn)練精確度隨著時(shí)間線性增長，直到接近100%,然而我們的驗(yàn)證精度卻停在96%~97%。 我們的驗(yàn)證損失在第4個(gè)循環(huán)(epochs)之后達(dá)到最小值，然后停頓，而訓(xùn)練損失在線性上保持下降直到接近０。 5.使用數(shù)據(jù)填充
數(shù)據(jù)增加采用從現(xiàn)有訓(xùn)練樣本生成更多訓(xùn)練數(shù)據(jù)的方法，通過產(chǎn)生可信的圖像的多個(gè)隨機(jī)變換來"增加"樣本。目標(biāo)是在訓(xùn)練的時(shí)候，我們的模型永遠(yuǎn)不會再看到完全相同的畫面兩次。這有助于模型學(xué)習(xí)到數(shù)據(jù)的更多方面，并更好地推廣。 在keras中，可以通過配置對ImageDataGenerator實(shí)例讀取的圖像執(zhí)行多個(gè)隨機(jī)變換來完成
datagen
= ImageDataGenerator
( rotation_range
= 40 , width_shift_range
= 0.2 , height_shift_range
= 0.2 , shear_range
= 0.2 , zoom_range
= 0.2 , horizontal_flip
= True , fill_mode
= 'nearest' )
這里列出一些可用的選項(xiàng)(更多選項(xiàng)，可以參考keras文檔），快速看一下這些參數(shù): rotation_range 是以度（0-180)為單位的值，它是隨機(jī)旋轉(zhuǎn)圖片的范圍 width_shift 和 height_shift 是范圍(占總寬度或高度的一小部分),用于縱向或橫向隨機(jī)轉(zhuǎn)換圖片 shear_range 用于隨機(jī)剪切變換 zoom_range 用于隨機(jī)放大圖片內(nèi)容 horizontal_flip 用于在沒有水平不對稱假設(shè)（例如真實(shí)世界圖片）的情況下水平地隨機(jī)翻轉(zhuǎn)一半圖像 fill_mode 是用于填充新創(chuàng)建的像素的策略，可以在旋轉(zhuǎn)或?qū)?高移位后顯示 import matplotlib
. pyplot
as plt
from keras
. preprocessing
import image
fnames
= [ os
. path
. join
( train_havemask_dir
, fname
) for fname
in os
. listdir
( train_havemask_dir
) ]
img_path
= fnames
[ 3 ]
img
= image
. load_img
( img_path
, target_size
= ( 150 , 150 ) )
x
= image
. img_to_array
( img
)
x
= x
. reshape
( ( 1 , ) + x
. shape
)
i
= 0
for batch
in datagen
. flow
( x
, batch_size
= 1 ) : plt
. figure
( i
) imgplot
= plt
. imshow
( image
. array_to_img
( batch
[ 0 ] ) ) i
+= 1 if i
% 4 == 0 : break
plt
. show
( )
如果我們使用這種數(shù)據(jù)增強(qiáng)配置來訓(xùn)練一個(gè)新的網(wǎng)絡(luò)，我們的網(wǎng)絡(luò)將永遠(yuǎn)不會看到相同重復(fù)的輸入。 然而，它看到的輸入仍然是相互關(guān)聯(lián)的，因?yàn)樗鼈儊碜陨倭康脑紙D像 - 我們不能產(chǎn)生新的信息，我們只能重新混合現(xiàn)有的信息。 因此，這可能不足以完全擺脫過度擬合(overfitting)。為了進(jìn)一步克服過度擬合(overfitting)，我們還將在密集連接(densely-connected)的分類器之前添加一個(gè)Dropout層。
model
= models
. Sequential
( )
model
. add
( layers
. Conv2D
( 32 , ( 3 , 3 ) , activation
= 'relu' , input_shape
= ( 150 , 150 , 3 ) ) )
model
. add
( layers
. MaxPooling2D
( ( 2 , 2 ) ) )
model
. add
( layers
. Conv2D
( 64 , ( 3 , 3 ) , activation
= 'relu' ) )
model
. add
( layers
. MaxPooling2D
( ( 2 , 2 ) ) )
model
. add
( layers
. Conv2D
( 128 , ( 3 , 3 ) , activation
= 'relu' ) )
model
. add
( layers
. MaxPooling2D
( ( 2 , 2 ) ) )
model
. add
( layers
. Conv2D
( 128 , ( 3 , 3 ) , activation
= 'relu' ) )
model
. add
( layers
. MaxPooling2D
( ( 2 , 2 ) ) )
model
. add
( layers
. Flatten
( ) )
model
. add
( layers
. Dropout
( 0.5 ) )
model
. add
( layers
. Dense
( 512 , activation
= 'relu' ) )
model
. add
( layers
. Dense
( 1 , activation
= 'sigmoid' ) ) model
. compile ( loss
= 'binary_crossentropy' , optimizer
= optimizers
. RMSprop
( lr
= 1e - 4 ) , metrics
= [ 'acc' ] )
使用數(shù)據(jù)填充(data augmentation)和dropout來訓(xùn)練我們的網(wǎng)絡(luò)
train_datagen
= ImageDataGenerator
( rescale
= 1 . / 255 , rotation_range
= 40 , width_shift_range
= 0.2 , height_shift_range
= 0.2 , shear_range
= 0.2 , zoom_range
= 0.2 , horizontal_flip
= True , )
test_datagen
= ImageDataGenerator
( rescale
= 1 . / 255 ) train_generator
= train_datagen
. flow_from_directory
( train_dir
, target_size
= ( 150 , 150 ) , batch_size
= 32 , class_mode
= 'binary' ) validation_generator
= test_datagen
. flow_from_directory
( validation_dir
, target_size
= ( 150 , 150 ) , batch_size
= 32 , class_mode
= 'binary' ) history
= model
. fit_generator
( train_generator
, steps_per_epoch
= 100 , epochs
= 60 , validation_data
= validation_generator
, validation_steps
= 50 )
Found 600 images belonging to 2 classes.
Found 300 images belonging to 2 classes.
Epoch 1/60
100/100 [==============================] - 351s 4s/step - loss: 0.4850 - acc: 0.7632 - val_loss: 0.2380 - val_acc: 0.8900
Epoch 2/60
100/100 [==============================] - 323s 3s/step - loss: 0.3041 - acc: 0.8703 - val_loss: 0.2513 - val_acc: 0.8833
Epoch 3/60
100/100 [==============================] - 322s 3s/step - loss: 0.2864 - acc: 0.8725 - val_loss: 0.2486 - val_acc: 0.8867
Epoch 4/60
100/100 [==============================] - 316s 3s/step - loss: 0.2490 - acc: 0.8964 - val_loss: 0.1243 - val_acc: 0.9533
Epoch 5/60
100/100 [==============================] - 308s 3s/step - loss: 0.2303 - acc: 0.9056 - val_loss: 0.1830 - val_acc: 0.9200
Epoch 6/60
100/100 [==============================] - 306s 3s/step - loss: 0.2098 - acc: 0.9152 - val_loss: 0.1101 - val_acc: 0.9633
Epoch 7/60
100/100 [==============================] - 312s 3s/step - loss: 0.1905 - acc: 0.9200 - val_loss: 0.1417 - val_acc: 0.9367
Epoch 8/60
100/100 [==============================] - 310s 3s/step - loss: 0.1771 - acc: 0.9272 - val_loss: 0.1021 - val_acc: 0.9700
Epoch 9/60
100/100 [==============================] - 302s 3s/step - loss: 0.1710 - acc: 0.9284 - val_loss: 0.1220 - val_acc: 0.9467
Epoch 10/60
100/100 [==============================] - 321s 3s/step - loss: 0.1618 - acc: 0.9375 - val_loss: 0.0920 - val_acc: 0.9667
Epoch 11/60
100/100 [==============================] - 308s 3s/step - loss: 0.1458 - acc: 0.9420 - val_loss: 0.2019 - val_acc: 0.9167
Epoch 12/60
100/100 [==============================] - 303s 3s/step - loss: 0.1411 - acc: 0.9456 - val_loss: 0.0829 - val_acc: 0.9700
Epoch 13/60
100/100 [==============================] - 304s 3s/step - loss: 0.1190 - acc: 0.9537 - val_loss: 0.0932 - val_acc: 0.9667
Epoch 14/60
100/100 [==============================] - 307s 3s/step - loss: 0.1163 - acc: 0.9569 - val_loss: 0.1085 - val_acc: 0.9567
Epoch 15/60
100/100 [==============================] - 306s 3s/step - loss: 0.1006 - acc: 0.9629 - val_loss: 0.0715 - val_acc: 0.9767
Epoch 16/60
100/100 [==============================] - 312s 3s/step - loss: 0.0960 - acc: 0.9667 - val_loss: 0.0588 - val_acc: 0.9767
Epoch 17/60
100/100 [==============================] - 308s 3s/step - loss: 0.0806 - acc: 0.9676 - val_loss: 0.0535 - val_acc: 0.9800
Epoch 18/60
100/100 [==============================] - 305s 3s/step - loss: 0.0778 - acc: 0.9711 - val_loss: 0.2239 - val_acc: 0.9300
Epoch 19/60
100/100 [==============================] - 307s 3s/step - loss: 0.0761 - acc: 0.9713 - val_loss: 0.0575 - val_acc: 0.9767
Epoch 20/60
100/100 [==============================] - 308s 3s/step - loss: 0.0507 - acc: 0.9816 - val_loss: 0.0926 - val_acc: 0.9667
Epoch 21/60
100/100 [==============================] - 306s 3s/step - loss: 0.0635 - acc: 0.9799 - val_loss: 0.0470 - val_acc: 0.9833
Epoch 22/60
100/100 [==============================] - 319s 3s/step - loss: 0.0701 - acc: 0.9750 - val_loss: 0.0437 - val_acc: 0.9867
Epoch 23/60
100/100 [==============================] - 315s 3s/step - loss: 0.0493 - acc: 0.9849 - val_loss: 0.0408 - val_acc: 0.9900
Epoch 24/60
100/100 [==============================] - 309s 3s/step - loss: 0.0513 - acc: 0.9824 - val_loss: 0.0449 - val_acc: 0.9767
Epoch 25/60
100/100 [==============================] - 304s 3s/step - loss: 0.0580 - acc: 0.9816 - val_loss: 0.0330 - val_acc: 0.9900
Epoch 26/60
100/100 [==============================] - 312s 3s/step - loss: 0.0434 - acc: 0.9884 - val_loss: 0.0357 - val_acc: 0.9833
Epoch 27/60
100/100 [==============================] - 302s 3s/step - loss: 0.0707 - acc: 0.9785 - val_loss: 0.0214 - val_acc: 0.9933
Epoch 28/60
100/100 [==============================] - 311s 3s/step - loss: 0.0431 - acc: 0.9869 - val_loss: 0.0306 - val_acc: 0.9900
Epoch 29/60
100/100 [==============================] - 305s 3s/step - loss: 0.0424 - acc: 0.9859 - val_loss: 0.0278 - val_acc: 0.9900
Epoch 30/60
100/100 [==============================] - 305s 3s/step - loss: 0.0240 - acc: 0.9934 - val_loss: 0.0233 - val_acc: 0.9933
Epoch 31/60
100/100 [==============================] - 335s 3s/step - loss: 0.0515 - acc: 0.9853 - val_loss: 0.0268 - val_acc: 0.9867
Epoch 32/60
100/100 [==============================] - 326s 3s/step - loss: 0.0515 - acc: 0.9884 - val_loss: 0.0222 - val_acc: 0.9933
Epoch 33/60
100/100 [==============================] - 320s 3s/step - loss: 0.0273 - acc: 0.9927 - val_loss: 0.0281 - val_acc: 0.9900
Epoch 34/60
100/100 [==============================] - 310s 3s/step - loss: 0.0411 - acc: 0.9909 - val_loss: 0.0282 - val_acc: 0.9900
Epoch 35/60
100/100 [==============================] - 306s 3s/step - loss: 0.0204 - acc: 0.9950 - val_loss: 0.0165 - val_acc: 0.9933
Epoch 36/60
100/100 [==============================] - 306s 3s/step - loss: 0.0623 - acc: 0.9842 - val_loss: 0.0268 - val_acc: 0.9900
Epoch 37/60
100/100 [==============================] - 304s 3s/step - loss: 0.0325 - acc: 0.9908 - val_loss: 0.0152 - val_acc: 0.9933
Epoch 38/60
100/100 [==============================] - 305s 3s/step - loss: 0.0178 - acc: 0.9933 - val_loss: 0.0117 - val_acc: 0.9967
Epoch 39/60
100/100 [==============================] - 309s 3s/step - loss: 0.0507 - acc: 0.9884 - val_loss: 0.0164 - val_acc: 0.9933
Epoch 40/60
100/100 [==============================] - 305s 3s/step - loss: 0.0398 - acc: 0.9919 - val_loss: 0.0236 - val_acc: 0.9933
Epoch 41/60
100/100 [==============================] - 300s 3s/step - loss: 0.0243 - acc: 0.9909 - val_loss: 0.0176 - val_acc: 0.9933
Epoch 42/60
100/100 [==============================] - 307s 3s/step - loss: 0.0419 - acc: 0.9922 - val_loss: 0.0145 - val_acc: 0.9933
Epoch 43/60
100/100 [==============================] - 302s 3s/step - loss: 0.0451 - acc: 0.9928 - val_loss: 0.0155 - val_acc: 0.9933
Epoch 44/60
100/100 [==============================] - 304s 3s/step - loss: 0.0640 - acc: 0.9893 - val_loss: 0.2175 - val_acc: 0.9333
Epoch 45/60
100/100 [==============================] - 314s 3s/step - loss: 0.0285 - acc: 0.9934 - val_loss: 0.0092 - val_acc: 0.9967
Epoch 46/60
100/100 [==============================] - 309s 3s/step - loss: 0.0279 - acc: 0.9937 - val_loss: 0.0116 - val_acc: 0.9933
Epoch 47/60
100/100 [==============================] - 305s 3s/step - loss: 0.0258 - acc: 0.9925 - val_loss: 0.0157 - val_acc: 0.9900
Epoch 48/60
100/100 [==============================] - 307s 3s/step - loss: 0.0319 - acc: 0.9906 - val_loss: 0.0142 - val_acc: 0.9933
Epoch 49/60
100/100 [==============================] - 305s 3s/step - loss: 0.0562 - acc: 0.9884 - val_loss: 0.0228 - val_acc: 0.9933
Epoch 50/60
100/100 [==============================] - 305s 3s/step - loss: 0.0370 - acc: 0.9931 - val_loss: 0.0230 - val_acc: 0.9867
Epoch 51/60
100/100 [==============================] - 309s 3s/step - loss: 0.0047 - acc: 0.9984 - val_loss: 0.0147 - val_acc: 0.9933
Epoch 52/60
100/100 [==============================] - 306s 3s/step - loss: 0.0237 - acc: 0.9941 - val_loss: 0.0161 - val_acc: 0.9900
Epoch 53/60
100/100 [==============================] - 301s 3s/step - loss: 0.0278 - acc: 0.9950 - val_loss: 0.0202 - val_acc: 0.9933
Epoch 54/60
100/100 [==============================] - 309s 3s/step - loss: 0.0266 - acc: 0.9945 - val_loss: 0.0267 - val_acc: 0.9933
Epoch 55/60
100/100 [==============================] - 302s 3s/step - loss: 0.0264 - acc: 0.9941 - val_loss: 0.0231 - val_acc: 0.9967
Epoch 56/60
100/100 [==============================] - 304s 3s/step - loss: 0.0132 - acc: 0.9959 - val_loss: 0.0177 - val_acc: 0.9933
Epoch 57/60
100/100 [==============================] - 326s 3s/step - loss: 0.0773 - acc: 0.9891 - val_loss: 0.0893 - val_acc: 0.9733
Epoch 58/60
100/100 [==============================] - 311s 3s/step - loss: 0.0049 - acc: 0.9984 - val_loss: 0.0277 - val_acc: 0.9933
Epoch 59/60
100/100 [==============================] - 308s 3s/step - loss: 0.0791 - acc: 0.9906 - val_loss: 0.0314 - val_acc: 0.9867
Epoch 60/60
100/100 [==============================] - 307s 3s/step - loss: 0.0133 - acc: 0.9956 - val_loss: 0.0186 - val_acc: 0.9933
查看0與1代表含義（0代表有口罩、1代表沒有口罩）
train_generator
. class_indices
{'have_mask': 0, 'no_mask': 1}
model
. save
( 'maskout/maskAndNomask_2.h5' )
保存模型后的文件夾顯示效果 繪制數(shù)據(jù)增強(qiáng)后的訓(xùn)練集與驗(yàn)證集的精確度與損失度的圖形，看一遍結(jié)果
acc
= history
. history
[ 'acc' ]
val_acc
= history
. history
[ 'val_acc' ]
loss
= history
. history
[ 'loss' ]
val_loss
= history
. history
[ 'val_loss' ] epochs
= range ( len ( acc
) ) plt
. plot
( epochs
, acc
, 'bo' , label
= 'Training acc' )
plt
. plot
( epochs
, val_acc
, 'b' , label
= 'Validation acc' )
plt
. title
( 'Training and validation accuracy' )
plt
. legend
( ) plt
. figure
( ) plt
. plot
( epochs
, loss
, 'bo' , label
= 'Training loss' )
plt
. plot
( epochs
, val_loss
, 'b' , label
= 'Validation loss' )
plt
. title
( 'Training and validation loss' )
plt
. legend
( ) plt
. show
( )
由于數(shù)據(jù)增加(data augmentation)和dropout使用，不再有過度擬合(overfitting)的問題。 訓(xùn)練曲線相當(dāng)密切地跟隨著驗(yàn)證曲線。訓(xùn)練精度和驗(yàn)證精度經(jīng)過60個(gè)循環(huán)無限接近100%。 驗(yàn)證損失和訓(xùn)練損失在線性上保持下降直到接近０。 通過進(jìn)一步利用正規(guī)化技術(shù)，及調(diào)整網(wǎng)絡(luò)參數(shù)(例如每個(gè)卷積層的濾波器數(shù)量或網(wǎng)絡(luò)層數(shù))，可以獲得更好的準(zhǔn)確度。
總結(jié)體會：可以看出，采用卷積神經(jīng)網(wǎng)絡(luò)（CNN）對人臉微笑識別和口罩識別出來的準(zhǔn)確的還是蠻高的。由于寒假忘記帶電腦回家（沒想到有疫情），用家里面的電腦運(yùn)行性能不太好，所有導(dǎo)致我訓(xùn)練時(shí)花了很多時(shí)間，但我一般都是晚上訓(xùn)練，第二天早上起來一般就會有結(jié)果了。
6.對人臉戴口罩與否的模型訓(xùn)練
判斷的第一張圖片（D:/mango/nana.jpg路徑下） from keras
. preprocessing
import image
from keras
. models
import load_model
import numpy
as npmodel
= load_model
( 'maskout/maskAndNomask_1.h5' ) img_path
= 'D:/mango/nana.jpg' img
= image
. load_img
( img_path
, target_size
= ( 150 , 150 ) )
img_tensor
= image
. img_to_array
( img
) / 255.0
img_tensor
= np
. expand_dims
( img_tensor
, axis
= 0 )
prediction
= model
. predict
( img_tensor
)
print ( prediction
)
if prediction
[ 0 ] [ 0 ] > 0.5 : result
= '未戴口罩'
else : result
= '戴口罩'
print ( result
)
[[ 0.0132275]]
戴口罩
判斷的第二張圖片（D:/mango/mengmeng.jpg路徑下） from keras
. preprocessing
import image
from keras
. models
import load_model
import numpy
as npmodel
= load_model
( 'maskout/maskAndNomask_2.h5' ) img_path
= 'D:/mango/mengmeng.jpg' img
= image
. load_img
( img_path
, target_size
= ( 150 , 150 ) )
img_tensor
= image
. img_to_array
( img
) / 255.0
img_tensor
= np
. expand_dims
( img_tensor
, axis
= 0 )
prediction
= model
. predict
( img_tensor
)
print ( prediction
)
if prediction
[ 0 ] [ 0 ] > 0.5 : result
= '未戴口罩'
else : result
= '戴口罩'
print ( result
)
[[ 0.99999881]]
未戴口罩
可以看見，判斷圖片是否戴口罩準(zhǔn)確度還是很高的，但是還是有一定得誤差。 綜上所述，圖片中人臉越清晰越容易判別正確的精度就越高。 四、完成一個(gè)攝像頭采集自己人臉、并對表情（笑臉/非笑臉、戴口罩和沒戴口罩）的實(shí)時(shí)分類判讀（輸出分類文字）的程序
1.笑臉/非笑臉實(shí)時(shí)分類判讀（輸出分類文字）的程序
import cv2
from keras
. preprocessing
import image
from keras
. models
import load_model
import numpy
as np
import dlib
from PIL
import Image
model
= load_model
( 'mangoout/smileAndUnsmile_2.h5' )
detector
= dlib
. get_frontal_face_detector
( )
video
= cv2
. VideoCapture
( 0 )
font
= cv2
. FONT_HERSHEY_SIMPLEX
def rec ( img
) : gray
= cv2
. cvtColor
( img
, cv2
. COLOR_BGR2GRAY
) dets
= detector
( gray
, 1 ) if dets
is not None : for face
in dets
: left
= face
. left
( ) top
= face
. top
( ) right
= face
. right
( ) bottom
= face
. bottom
( ) cv2
. rectangle
( img
, ( left
, top
) , ( right
, bottom
) , ( 0 , 255 , 0 ) , 2 ) img1
= cv2
. resize
( img
[ top
: bottom
, left
: right
] , dsize
= ( 150 , 150 ) ) img1
= cv2
. cvtColor
( img1
, cv2
. COLOR_BGR2RGB
) img1
= np
. array
( img1
) / 255 . img_tensor
= img1
. reshape
( - 1 , 150 , 150 , 3 ) prediction
= model
. predict
( img_tensor
) if prediction
[ 0 ] [ 0 ] > 0.5 : result
= 'unsmile' else : result
= 'smile' cv2
. putText
( img
, result
, ( left
, top
) , font
, 2 , ( 0 , 255 , 0 ) , 2 , cv2
. LINE_AA
) cv2
. imshow
( 'Video' , img
)
while video
. isOpened
( ) : res
, img_rd
= video
. read
( ) if not res
: break rec
( img_rd
) if cv2
. waitKey
( 1 ) & 0xFF == ord ( 'q' ) : break
video
. release
( )
cv2
. destroyAllWindows
( )
2.戴口罩和沒戴口罩的實(shí)時(shí)分類判讀（輸出分類文字）的程序
import cv2
from keras
. preprocessing
import image
from keras
. models
import load_model
import numpy
as np
import dlib
from PIL
import Image
model
= load_model
( 'maskout/maskAndNomask_2.h5' )
detector
= dlib
. get_frontal_face_detector
( )
video
= cv2
. VideoCapture
( 0 )
font
= cv2
. FONT_HERSHEY_SIMPLEX
def rec ( img
) : gray
= cv2
. cvtColor
( img
, cv2
. COLOR_BGR2GRAY
) dets
= detector
( gray
, 1 ) if dets
is not None : for face
in dets
: left
= face
. left
( ) top
= face
. top
( ) right
= face
. right
( ) bottom
= face
. bottom
( ) cv2
. rectangle
( img
, ( left
, top
) , ( right
, bottom
) , ( 0 , 255 , 255 ) , 2 )
def mask ( img
) : img1
= cv2
. resize
( img
, dsize
= ( 150 , 150 ) ) img1
= cv2
. cvtColor
( img1
, cv2
. COLOR_BGR2RGB
) img1
= np
. array
( img1
) / 255 . img_tensor
= img1
. reshape
( - 1 , 150 , 150 , 3 ) prediction
= model
. predict
( img_tensor
) if prediction
[ 0 ] [ 0 ] > 0.5 : result
= 'no-mask' else : result
= 'have-mask' cv2
. putText
( img
, result
, ( 100 , 200 ) , font
, 2 , ( 255 , 255 , 100 ) , 2 , cv2
. LINE_AA
) cv2
. imshow
( 'Video' , img
)
while video
. isOpened
( ) : res
, img_rd
= video
. read
( ) if not res
: break rec
( img_rd
) mask
( img_rd
) if cv2
. waitKey
( 1 ) & 0xFF == ord ( 'q' ) : break
video
. release
( )
cv2
. destroyAllWindows
( )
這次的關(guān)于用Python編碼，利用卷積神經(jīng)網(wǎng)路（CNN）實(shí)現(xiàn)的笑臉識別和口罩識別到這里就結(jié)束了。這次把自己圖片也用上了，主要是想讓自己寫的博客盡量有意義一點(diǎn)（平時(shí)生活中拍照還是用美圖好一點(diǎn)，人丑勿怪，哈哈… …）。
總結(jié)
以上是生活随笔 為你收集整理的【人工智能与机器学习】——Keras编程分别实现人脸微笑和口罩数据集的识别模型训练和测试（卷积神经网络CNN） + 实时分类微笑和口罩识别检测 的全部內(nèi)容，希望文章能夠幫你解決所遇到的問題。
如果覺得生活随笔 網(wǎng)站內(nèi)容還不錯(cuò)，歡迎將生活随笔 推薦給好友。
