画像認識プログラムが学習していない原因

Question

手書き画像画像の数字の認識のプログラムを作っています。
しかし、損失が下がらず、認識率も低いまま収束してしまいます。
損失を見た感じ学習をしていない感じです。原因がわからないのでわかる方いましたらご指摘お願いいたします。

発生している問題

Train on 160 samples, validate on 40 samples
Epoch 1/30
160/160 [==============================] - 3s 21ms/step - loss: 2.3505 - accuracy: 0.0875 - val_loss: 2.2978 - val_accuracy: 0.1500
Epoch 2/30
160/160 [==============================] - 0s 2ms/step - loss: 2.3065 - accuracy: 0.0875 - val_loss: 2.3016 - val_accuracy: 0.0500
Epoch 3/30
160/160 [==============================] - 0s 2ms/step - loss: 2.2994 - accuracy: 0.1250 - val_loss: 2.3053 - val_accuracy: 0.0500
Epoch 4/30
160/160 [==============================] - 0s 2ms/step - loss: 2.2917 - accuracy: 0.1250 - val_loss: 2.3207 - val_accuracy: 0.0500
Epoch 5/30
160/160 [==============================] - 0s 2ms/step - loss: 2.2993 - accuracy: 0.1250 - val_loss: 2.3048 - val_accuracy: 0.0500
Epoch 6/30
160/160 [==============================] - 0s 2ms/step - loss: 2.2917 - accuracy: 0.1250 - val_loss: 2.2992 - val_accuracy: 0.0500
Epoch 7/30
160/160 [==============================] - 0s 2ms/step - loss: 2.2929 - accuracy: 0.1250 - val_loss: 2.2986 - val_accuracy: 0.0500
Epoch 8/30
160/160 [==============================] - 0s 2ms/step - loss: 2.2885 - accuracy: 0.1250 - val_loss: 2.2991 - val_accuracy: 0.0500
Epoch 9/30
160/160 [==============================] - 0s 2ms/step - loss: 2.2877 - accuracy: 0.1250 - val_loss: 2.3033 - val_accuracy: 0.0500
Epoch 10/30
160/160 [==============================] - 0s 2ms/step - loss: 2.2881 - accuracy: 0.1250 - val_loss: 2.3010 - val_accuracy: 0.0500
Epoch 11/30
160/160 [==============================] - 0s 2ms/step - loss: 2.2890 - accuracy: 0.1250 - val_loss: 2.3040 - val_accuracy: 0.0500
Epoch 12/30
160/160 [==============================] - 0s 2ms/step - loss: 2.2867 - accuracy: 0.1250 - val_loss: 2.2970 - val_accuracy: 0.0500
Epoch 13/30
160/160 [==============================] - 0s 2ms/step - loss: 2.2861 - accuracy: 0.1250 - val_loss: 2.2954 - val_accuracy: 0.0500
Epoch 14/30
160/160 [==============================] - 0s 2ms/step - loss: 2.2862 - accuracy: 0.1250 - val_loss: 2.2940 - val_accuracy: 0.0500
Epoch 15/30
160/160 [==============================] - 0s 2ms/step - loss: 2.2904 - accuracy: 0.1250 - val_loss: 2.2979 - val_accuracy: 0.0500
Epoch 16/30
160/160 [==============================] - 0s 2ms/step - loss: 2.2841 - accuracy: 0.1250 - val_loss: 2.2940 - val_accuracy: 0.0500
Epoch 17/30
160/160 [==============================] - 0s 2ms/step - loss: 2.2855 - accuracy: 0.1312 - val_loss: 2.2938 - val_accuracy: 0.1250
Epoch 18/30
160/160 [==============================] - 0s 2ms/step - loss: 2.2862 - accuracy: 0.1250 - val_loss: 2.2936 - val_accuracy: 0.1250
Epoch 19/30
160/160 [==============================] - 0s 2ms/step - loss: 2.2888 - accuracy: 0.1312 - val_loss: 2.2957 - val_accuracy: 0.1250
Epoch 20/30
160/160 [==============================] - 0s 2ms/step - loss: 2.2861 - accuracy: 0.1312 - val_loss: 2.2975 - val_accuracy: 0.1250
Epoch 21/30
160/160 [==============================] - 0s 2ms/step - loss: 2.2871 - accuracy: 0.1187 - val_loss: 2.2974 - val_accuracy: 0.0500
Epoch 22/30
160/160 [==============================] - 0s 2ms/step - loss: 2.2859 - accuracy: 0.1250 - val_loss: 2.2949 - val_accuracy: 0.0500
Epoch 23/30
160/160 [==============================] - 0s 2ms/step - loss: 2.2853 - accuracy: 0.1312 - val_loss: 2.2952 - val_accuracy: 0.0500
Epoch 24/30
160/160 [==============================] - 0s 2ms/step - loss: 2.2856 - accuracy: 0.1063 - val_loss: 2.2953 - val_accuracy: 0.0500
Epoch 25/30
160/160 [==============================] - 0s 2ms/step - loss: 2.2850 - accuracy: 0.0875 - val_loss: 2.2968 - val_accuracy: 0.0500
Epoch 26/30
160/160 [==============================] - 0s 2ms/step - loss: 2.2849 - accuracy: 0.1250 - val_loss: 2.3000 - val_accuracy: 0.0500
Epoch 27/30
160/160 [==============================] - 0s 2ms/step - loss: 2.2854 - accuracy: 0.1250 - val_loss: 2.2984 - val_accuracy: 0.0500
Epoch 28/30
160/160 [==============================] - 0s 2ms/step - loss: 2.2855 - accuracy: 0.1375 - val_loss: 2.2958 - val_accuracy: 0.1250
Epoch 29/30
160/160 [==============================] - 0s 2ms/step - loss: 2.2843 - accuracy: 0.1312 - val_loss: 2.2929 - val_accuracy: 0.1250
Epoch 30/30
160/160 [==============================] - 0s 2ms/step - loss: 2.2853 - accuracy: 0.1312 - val_loss: 2.2921 - val_accuracy: 0.1250

該当のソースコード

import numpy as np
import matplotlib.pyplot as plt
import glob,os
from sklearn.model_selection import train_test_split
import cv2
from keras.layers import Convolution2D, BatchNormalization, Activation, MaxPooling2D, Add, Dropout,  Flatten, Dense
from keras import optimizers
from keras.utils import to_categorical

from keras import models
from keras import layers

x=[]
z=[]

#画像フォルダのパスを指定
path=os.path.dirname(os.path.abspath('__file__'))
in_size=(28,28)
out_size=10

path_one  =  path +'/Dataset/one'
path_two  =  path +'/Dataset/two'
path_three=  path +'/Dataset/three'
path_four =  path +'/Dataset/four'


file_1=glob.glob(path_one +'/*.jpg')
file_2=glob.glob(path_two +'/*.jpg')
file_3=glob.glob(path_three +'/*.jpg')
file_4=glob.glob(path_four +'/*.jpg')


print(len(file_1))
print(len(file_2))
print(len(file_3))
print(len(file_4))
print(len(file_5))
print(len(file_6))
print(len(file_7))
print(len(file_8))
print(len(file_9))
print(len(file_0))

#出力
26
20
23
21
11
19
20
16
22
22




def load_dir_1(path,label):

    for i in file_1:
        img=cv2.imread(i)
        img=cv2.cvtColor(img,cv2.COLOR_BGR2GRAY)
        img=cv2.resize(img,in_size)
        img=img/255.0
        x.append(img)
        z.append(label)

    return [x,z]


def load_dir_2(path,label):

    for i in file_2:
        img=cv2.imread(i)
        img=cv2.cvtColor(img,cv2.COLOR_BGR2GRAY)
        img=cv2.resize(img,in_size)
        img=img/255.0
        x.append(img)
        z.append(label)
    return [x,z]

def load_dir_3(path,label):

    for i in file_3:
        img=cv2.imread(i)
        img=cv2.cvtColor(img,cv2.COLOR_BGR2GRAY)
        img=cv2.resize(img,in_size)
        img=img/255.0
        x.append(img)
        z.append(label)
    return [x,z]

def load_dir_4(path,label):

    for i in file_4:
        img=cv2.imread(i)
        img=cv2.cvtColor(img,cv2.COLOR_BGR2GRAY)
        img=cv2.resize(img,in_size)
        img=img/255.0
        x.append(img)
        z.append(label)
    return [x,z]




load_dir_1(path_one,1)

load_dir_2(path_two,2)

load_dir_3(path_three,3)

load_dir_4(path_four,4)


x=np.array(x)
z=np.array(z)


import keras
x_train,x_test,z_train,z_test=train_test_split(x,z,test_size=0.2)

x_train=x_train.reshape(len(x_train),28,28,1).astype('float32')
x_test=x_test.reshape(len(x_test),28,28,1).astype('float32')


z_train=keras.utils.np_utils.to_categorical(z_train.astype('int32'),10)
z_test=keras.utils.np_utils.to_categorical(z_test.astype('int32'),10)

model=models.Sequential()
model.add(layers.Conv2D(32,(3,3),activation='relu',input_shape=(28,28,1)))
model.add(layers.MaxPooling2D((2,2)))
model.add(layers.Conv2D(64,(3,3),activation='relu'))
model.add(layers.MaxPooling2D(2,2))
model.add(Dropout(0.2))
model.add(layers.Conv2D(64,(3,3),activation='relu'))
model.add(layers.Flatten())
model.add(layers.Dense(64,activation='relu'))
model.add(layers.Dense(10,activation='softmax'))

from keras import optimizers
model.compile(optimizer='adam',
             loss='categorical_crossentropy',
             metrics=['accuracy'])
history = model.fit(x_train, z_train, epochs=30, batch_size=32,validation_data=(x_test, z_test))

Comments

[can110]

いくつか未定義の変数があるようです。エラーが発生せず動作するコードを提示ください。

[koukiten]

修正しました

[can110]

「NameError: name 'path_one_speckle' is not defined」にて動作しません。
質問に記載しているコードを「そのままコピペ」して実行してみてください（動作しますか？）

[koukiten]

何度もすみません。
関数に入れるパスを間違えていたので直しました。

[can110]

「x_train=x_train.reshape(len(x_train),28,28,1).astype('float32')」にて
「AttributeError: 'list' object has no attribute 'reshape'」が発生します。
質問に記載しているコードを「そのままコピペ」して実行してみてください（動作しますか？）

[koukiten]

本当にすみません！
リストを配列に変えるのを忘れてたので修正しました。

[can110]

とりあえず、修正された質問に記載しているコードを「そのままコピペ」して実行してみてください。
動作しますか？

[koukiten]

必要なライブラリのインポートを修正しました。
動作しました。

[can110]

入力画像に問題がありそうなので、まずはoneからzeroまでの各ファイル数、len(file_1)～len(file_0)をprintした結果を提示ください。

[koukiten]

文字数制限を超えたので5以降の数字の読み込みの関数を削除しました。

Answer 1

このあたりの知識に乏しいので的確な原因は指摘できませんが、元画像データに問題があると思われます。
当方にて各数字の画像数１０個、各数字の画像は全く同一のもの（「１」なら１０個すべて同じカタチ「１」の画像）で動作させた結果です。
あたりまえですが、正解率は1.0に収束しています。

Train on 80 samples, validate on 20 samples
Epoch 1/30
80/80 [==============================] - 0s 2ms/step - loss: 2.3166 - accuracy: 0.1000 - val_loss: 2.3035 - val_accuracy: 0.0500
Epoch 2/30
80/80 [==============================] - 0s 411us/step - loss: 2.2934 - accuracy: 0.1125 - val_loss: 2.3140 - val_accuracy: 0.1500
Epoch 3/30
80/80 [==============================] - 0s 412us/step - loss: 2.2740 - accuracy: 0.1250 - val_loss: 2.3295 - val_accuracy: 0.1500
Epoch 4/30
80/80 [==============================] - 0s 387us/step - loss: 2.2501 - accuracy: 0.2000 - val_loss: 2.3635 - val_accuracy: 0.0500
Epoch 5/30
80/80 [==============================] - 0s 388us/step - loss: 2.2377 - accuracy: 0.1500 - val_loss: 2.4008 - val_accuracy: 0.0000e+00
Epoch 6/30
80/80 [==============================] - 0s 412us/step - loss: 2.2185 - accuracy: 0.1250 - val_loss: 2.3816 - val_accuracy: 0.0000e+00
Epoch 7/30
80/80 [==============================] - 0s 400us/step - loss: 2.1710 - accuracy: 0.1875 - val_loss: 2.3371 - val_accuracy: 0.1000
Epoch 8/30
80/80 [==============================] - 0s 400us/step - loss: 2.1129 - accuracy: 0.3875 - val_loss: 2.2777 - val_accuracy: 0.1500
Epoch 9/30
80/80 [==============================] - 0s 362us/step - loss: 2.0528 - accuracy: 0.5000 - val_loss: 2.2032 - val_accuracy: 0.2500
Epoch 10/30
80/80 [==============================] - 0s 387us/step - loss: 1.9785 - accuracy: 0.4125 - val_loss: 2.1106 - val_accuracy: 0.0500
Epoch 11/30
80/80 [==============================] - 0s 375us/step - loss: 1.8595 - accuracy: 0.4500 - val_loss: 1.9630 - val_accuracy: 0.2500
Epoch 12/30
80/80 [==============================] - 0s 350us/step - loss: 1.7170 - accuracy: 0.6000 - val_loss: 1.7560 - val_accuracy: 0.5000
Epoch 13/30
80/80 [==============================] - 0s 375us/step - loss: 1.5362 - accuracy: 0.7750 - val_loss: 1.4589 - val_accuracy: 1.0000
Epoch 14/30
80/80 [==============================] - 0s 375us/step - loss: 1.3673 - accuracy: 0.8625 - val_loss: 1.2133 - val_accuracy: 1.0000
Epoch 15/30
80/80 [==============================] - 0s 362us/step - loss: 1.1369 - accuracy: 0.8375 - val_loss: 0.9378 - val_accuracy: 1.0000
Epoch 16/30
80/80 [==============================] - 0s 363us/step - loss: 0.9292 - accuracy: 0.9125 - val_loss: 0.7440 - val_accuracy: 0.9000
Epoch 17/30
80/80 [==============================] - 0s 363us/step - loss: 0.7212 - accuracy: 0.9375 - val_loss: 0.5730 - val_accuracy: 1.0000
Epoch 18/30
80/80 [==============================] - 0s 363us/step - loss: 0.6000 - accuracy: 0.8875 - val_loss: 0.4320 - val_accuracy: 1.0000
Epoch 19/30
80/80 [==============================] - 0s 363us/step - loss: 0.4228 - accuracy: 0.9875 - val_loss: 0.3599 - val_accuracy: 0.9000
Epoch 20/30
80/80 [==============================] - 0s 363us/step - loss: 0.4129 - accuracy: 0.8625 - val_loss: 0.2648 - val_accuracy: 1.0000
Epoch 21/30
80/80 [==============================] - 0s 363us/step - loss: 0.3232 - accuracy: 0.9375 - val_loss: 0.2445 - val_accuracy: 0.9000
Epoch 22/30
80/80 [==============================] - 0s 375us/step - loss: 0.2665 - accuracy: 0.9375 - val_loss: 0.1590 - val_accuracy: 1.0000
Epoch 23/30
80/80 [==============================] - 0s 363us/step - loss: 0.2104 - accuracy: 0.9625 - val_loss: 0.1119 - val_accuracy: 1.0000
Epoch 24/30
80/80 [==============================] - 0s 375us/step - loss: 0.1193 - accuracy: 1.0000 - val_loss: 0.1258 - val_accuracy: 1.0000
Epoch 25/30
80/80 [==============================] - 0s 363us/step - loss: 0.1458 - accuracy: 0.9875 - val_loss: 0.0807 - val_accuracy: 1.0000
Epoch 26/30
80/80 [==============================] - 0s 350us/step - loss: 0.0805 - accuracy: 1.0000 - val_loss: 0.0575 - val_accuracy: 1.0000
Epoch 27/30
80/80 [==============================] - 0s 350us/step - loss: 0.0734 - accuracy: 1.0000 - val_loss: 0.0588 - val_accuracy: 1.0000
Epoch 28/30
80/80 [==============================] - 0s 338us/step - loss: 0.0585 - accuracy: 1.0000 - val_loss: 0.0338 - val_accuracy: 1.0000
Epoch 29/30
80/80 [==============================] - 0s 338us/step - loss: 0.0635 - accuracy: 1.0000 - val_loss: 0.0279 - val_accuracy: 1.0000
Epoch 30/30
80/80 [==============================] - 0s 350us/step - loss: 0.0384 - accuracy: 1.0000 - val_loss: 0.0375 - val_accuracy: 1.0000

Comments

[koukiten]

やはりそう考えるのが妥当ですよね。
画像の特徴量を見つけられていない感じがします。