-
Notifications
You must be signed in to change notification settings - Fork 0
/
image_recognition.py
399 lines (319 loc) · 14.3 KB
/
image_recognition.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
# -*- coding: utf-8 -*-
"""Image_Recognition.ipynb
Automatically generated by Colaboratory.
Original file is located at
https://colab.research.google.com/drive/127deDIjrYmegb-52HkkPSk53NCmwxsss
"""
!pip install --upgrade tensorflow
from __future__ import absolute_import, division, print_function, unicode_literals
# TensorFlow and tf.keras
import tensorflow as tf
from tensorflow import keras
print(tf.__version__)
from google.colab import drive
drive.mount('/content/drive', force_remount=True)
"""Home-made Classes and Functions"""
import matplotlib.pyplot as plt
import numpy as np
############ Classes ##############
class Data:
def __init__(self, image, label):
self.image = image
self.label = label
class _Deformer(object):
def getmesh(self, im):
x, y = im.size
return [(
# target rectangle
(0, 0, x, y),
# corresponding source quadrilateral (1)
# (NW, SW, SE, and NE. see method=QUAD)
(0, y, 0, 0, x, y, x, 0))]
############ Functions ##############
def show_images(images, cols = 1, titles = None):
"""Display a list of images in a single figure with matplotlib.
Parameters
---------
images: List of np.arrays compatible with plt.imshow.
cols (Default = 1): Number of columns in figure (number of rows is
set to np.ceil(n_images/float(cols))).
titles: List of titles corresponding to each image. Must have
the same length as titles.
"""
assert((titles is None)or (len(images) == len(titles)))
n_images = len(images)
if titles is None: titles = ['Image (%d)' % i for i in range(1,n_images + 1)]
fig = plt.figure()
for n, (image, title) in enumerate(zip(images, titles)):
a = fig.add_subplot(cols, np.ceil(n_images/float(cols)), n + 1)
if image.ndim == 2:
plt.gray()
plt.imshow(image)
a.set_title(title)
fig.set_size_inches(np.array(fig.get_size_inches()) * n_images)
plt.show()
"""Pré-traitements des images"""
import os
import numpy as np
from collections import Counter
import pandas as pd
styles = os.listdir("drive/My Drive/peintures")#liste les styles de peintures du dataset
occu = []
for style in styles:
peintures = os.listdir("drive/My Drive/peintures/"+style)#liste toutes les peintures d'un style particulier
for peinture in peintures:
occu.append(style)
styles_count = Counter(occu)#nombre de tableaux pour chaque style
df = pd.DataFrame.from_dict(styles_count, orient='index',columns=['Frequency'])
df = df.sort_values(by=['Frequency'])
df.plot.barh()#Affiche la repartition des peintures a travers les differents styles
from PIL import Image
from PIL import ImageOps
from PIL import ImageFilter
import matplotlib.pyplot as plt
images = []
new_images = []
im = Image.open("drive/My Drive/peintures/baroque/Caravaggio_30.jpg").convert('RGB').resize((150,150), Image.ANTIALIAS)
images.append(im)
images.append(ImageOps.flip(im))#on retourne l'image
images.append(ImageOps.mirror(im))#effet miroir
images.append(ImageOps.deform(im, deformer=_Deformer()))#distortion de l'image en utilisant un objet gitmesh
images.append(im.filter(ImageFilter.BLUR))#on floute l'image
for image in images:
new_images.append(np.array(image))
show_images(new_images,titles=['Original','Flip horizontal','Effet miroir','Distortion','Floutage'])
from PIL import Image
from PIL import ImageOps
from PIL import ImageFilter
import numpy as np
import os
import random
styles = os.listdir("drive/My Drive/peintures")#styles de peintures du dataset
dataset1 = []#images sans couleurs
dataset2 = []#images avec couleurs (RGB)
labels = []#reponse attendue pour l'image
for style in styles:
peintures = os.listdir("drive/My Drive/peintures/"+style)
peintures = random.choices(peintures, k=80)#choix au hasard de 20 peintures dans le repertoire
for peinture in peintures:
try:
file_name = "drive/My Drive/peintures/"+style+"/"+peinture
img2 = Image.open(file_name).convert('RGB').resize((150,150), Image.ANTIALIAS)#on formate l'image dans une taille standard et petit
img1 = ImageOps.grayscale(img2)
dataset1.append(np.array(img1)) # Ajout de l'image sans couleurs en format numpy.array dans dataset
dataset1.append(np.array(ImageOps.flip(img1)))
dataset1.append(np.array(ImageOps.mirror(img1)))
dataset1.append(np.array(ImageOps.deform(img1, deformer=_Deformer())))
dataset1.append(np.array(img1.filter(ImageFilter.BLUR)))
dataset2.append(np.array(img2)) # Ajout de l'image avec couleurs en format numpy.array dans dataset
dataset2.append(np.array(ImageOps.flip(img2)))
dataset2.append(np.array(ImageOps.mirror(img2)))
dataset2.append(np.array(ImageOps.deform(img2, deformer=_Deformer())))
dataset2.append(np.array(img2.filter(ImageFilter.BLUR)))
if(style=="baroque"):#ajout de la reponse attendue dans labels
labels.extend([0,0,0,0,0])
elif(style=="cubisme"):
labels.extend([1,1,1,1,1])
elif(style=="fauvisme"):
labels.extend([2,2,2,2,2])
elif(style=="gothique"):
labels.extend([3,3,3,3,3])
elif(style=="impressionisme"):
labels.extend([4,4,4,4,4])
elif(style=="moderne"):
labels.extend([5,5,5,5,5])
elif(style=="postimpressionisme"):
labels.extend([6,6,6,6,6])
elif(style=="realisme"):
labels.extend([7,7,7,7,7])
elif(style=="renaissance"):
labels.extend([8,8,8,8,8])
elif(style=="romantisme"):
labels.extend([9,9,9,9,9])
elif(style=="surrealisme"):
labels.extend([10,10,10,10,10])
elif(style=="symbolisme"):
labels.extend([11,11,11,11,11])
except OSError:
print("Can not open file " + file_name)
import numpy as np
FinalData1=[]
FinalData2=[]
for i in range (len(dataset1)):
FinalData1.append(Data(dataset1[i],labels[i]))
FinalData2.append(Data(dataset2[i],labels[i]))
"""Normalization des images"""
import numpy as np
#Normalization
#Inew = (I - I.mean) / I.std
for i in range (len(FinalData2)):
FinalData2[i].image = (FinalData2[i].image - np.mean(FinalData2[i].image))/np.std(FinalData2[i].image)
"""Visualisation finale d'un echantillon"""
import matplotlib.pyplot as plt
import numpy as np
#visualisation d'une image du dataset 1 et du dataset 2
print("###################### Dataset 1 (images sans couleurs) ######################")
show_images([FinalData1[20].image,FinalData1[21].image,FinalData1[22].image,FinalData1[23].image,FinalData1[24].image])
#print("###################### Dataset 2 (images avec couleurs) ######################")
#show_images([FinalData2[20].image,FinalData2[21].image,FinalData2[22].image,FinalData2[23].image,FinalData2[24].image])
"""RNN Classiques"""
#Construction du train set et du test set
from sklearn.model_selection import train_test_split
import numpy as np
train , test = train_test_split(FinalData1,test_size=0.25)
train_x = [data.image for data in train]
train_y = [data.label for data in train]
test_x = [data.image for data in test]
test_y = [data.label for data in test]
trainx=np.asarray(train_x)
testx=np.asarray(test_x)
trainy=np.asarray(train_y)
testy=np.asarray(test_y)
#On convertit les jeux de donnees en format tf.tensor pour que le format des donnees soit le plus adapte à notre RNN.
trainx = tf.convert_to_tensor(train_x, dtype=tf.float32)
testx = tf.convert_to_tensor(test_x, dtype=tf.float32)
import pandas as pd
import numpy as np
Results = pd.DataFrame([[2,0,0,0],[3,0,0,0],[4,0,0,0]],
index=["Model 1","Model 2","Model 3"],
columns=["nb hidden layers","nb weights","test loss","test acc"],
dtype=np.float128)
for i in range(0,3):
if(i==0):
#construction du model
model1 = tf.keras.Sequential([tf.keras.layers.Flatten(input_shape=(150,150)),
tf.keras.layers.Dense(256, activation='relu'),
tf.keras.layers.Dense(128, activation='elu'),
tf.keras.layers.Dense(12, activation='softmax')])
elif(i==1):
#construction du model
model1 = tf.keras.Sequential([tf.keras.layers.Flatten(input_shape=(150,150)),
tf.keras.layers.Dense(512, activation='relu'),
tf.keras.layers.Dense(256, activation='relu'),
tf.keras.layers.Dense(128, activation='elu'),
tf.keras.layers.Dense(12, activation='softmax')
])
else:
#construction du model
model1 = tf.keras.Sequential([
tf.keras.layers.Flatten(input_shape=(150,150)),
tf.keras.layers.Dense(784, activation='relu'),
tf.keras.layers.Dense(512, activation='relu'),
tf.keras.layers.Dense(256, activation='relu'),
tf.keras.layers.Dense(128, activation='elu'),
tf.keras.layers.Dense(12, activation='softmax')
])
#Definition des parametres d'apprentissage du RNN
model1.compile(optimizer='adam',loss='sparse_categorical_crossentropy',metrics=['accuracy'])
#Apprentissage
model1.fit(trainx, trainy, epochs=15)
#evaluation du modele
test_loss, test_acc = model1.evaluate(testx, testy, verbose=2)
Results["test loss"][i]=test_loss
Results["test acc"][i]=test_acc
Results["nb weights"][i]=model1.count_params()
del model1
print("______________________________________________________________________________________________")
print("\nResults\n\n")
Results
#On remarque qu'en augmentant le nombe de layers et donc le nombre de poids l'accuracy du RNN varie mais ne diminue pas forcement.
#Ici nous avons fixe le nombre de hidden layers et de neurones pour chacune de ces layers un peu au hasard.
#Il faudrait donc tester plein de models avec chacun un nombre de hidden layers et neurones different et esperer tomber sur la meilleure combinaison.
#Cette appproche n'est pas efficace et elle ne garantie même pas un resultat convenable. On va donc tester une autre approche.
"""RNN Convolutif"""
#Construction du train set et du test set
from sklearn.model_selection import train_test_split
import numpy as np
train2 , test2 = train_test_split(FinalData2,test_size=0.25)
train_x2 = [data.image for data in train2]
train_y2 = [data.label for data in train2]
test_x2 = [data.image for data in test2]
test_y2 = [data.label for data in test2]
trainx2=np.asarray(train_x2)
testx2=np.asarray(test_x2)
trainy2=np.asarray(train_y2)
testy2=np.asarray(test_y2)
#On convertit les jeux de donnees en format tf.tensor pour que le format des donnees soit le plus adapte à notre RNN.
trainx2 = tf.convert_to_tensor(trainx2, dtype=tf.float32)
testx2 = tf.convert_to_tensor(testx2, dtype=tf.float32)
import keras
from keras.models import Sequential
from keras.layers import Dense, Dropout, Conv2D, MaxPooling2D, BatchNormalization, Flatten
#construction d'un RNN convolutif
model2 = tf.keras.Sequential([
tf.keras.layers.Conv2D(32,(3,3),activation='relu',input_shape=(150,150,3)),
tf.keras.layers.MaxPooling2D(pool_size=(2,2)),
tf.keras.layers.BatchNormalization(),
tf.keras.layers.Conv2D(64,(3,3),activation='relu'),
tf.keras.layers.MaxPooling2D(pool_size=(2,2)),
tf.keras.layers.Flatten(),
tf.keras.layers.Dense(128, activation='relu'),
tf.keras.layers.Dense(12, activation='softmax')
])
#choix de l'algo d'apprentissage
model2.compile(optimizer='adam',loss='sparse_categorical_crossentropy',metrics=['accuracy'])
#Apprentissage du modele
model2.fit(trainx2, trainy2, epochs=7,steps_per_epoch=len(trainx2))
#evaluation du modele
test_loss2, test_acc2 = model2.evaluate(testx2, testy2, verbose=2)
print('\nTest accuracy:', test_acc2)
#Ici on se retrouve avec le meme probleme qu'un RNN classique. Il est necessaire de tester beaucoup de modeles en faisant varier le nombre de hidden layers,
#le nombre et le type de couche de convolution. On peut alors esperer tomber sur mla meilleure combianaison possible mais ce n'est âs garantie et la compilation de tous ces modeles
#prendrai enormement de temps. On va donc utiliser des modeles pre-entraines.
"""RNN with Transfer Learning **(VGG19)**"""
import keras
from keras.models import Sequential
from keras.layers import Dense, Dropout, Conv2D, MaxPooling2D, BatchNormalization, Flatten
#Import du modele VGG19 pre-entrainer sur le dataset inagenet
imagenet_model = tf.keras.applications.VGG19(weights = "imagenet",
include_top=False,
input_shape = (150, 150, 3),
pooling='max')
#Couches du modele VGG19
imagenet_model.layers
# freeze feature layers and rebuild model
for l in imagenet_model.layers:
l.trainable = False
#construction du model
model3 = [tf.keras.layers.Flatten(),
tf.keras.layers.Dense(128, activation='relu'),
tf.keras.layers.Dense(12, activation='softmax')]
#Jointure des deux modeles
model_using_pre_trained_one = tf.keras.Sequential( imagenet_model.layers + model3 )
#choix de l'algo d'apprentissage
model_using_pre_trained_one.compile(optimizer='adam',loss='sparse_categorical_crossentropy',metrics=['accuracy'])
#Apprentissage du modele
model_using_pre_trained_one.fit(trainx2, trainy2, epochs=7,steps_per_epoch=len(trainx2))
#evaluation du modele
test_loss3, test_acc3 = model_using_pre_trained_one.evaluate(testx2, testy2, verbose=2)
print('\nTest accuracy:', test_acc3)
print('\nTest loss:', test_loss3)
"""RNN with Transfer Learning **(VGG16)**"""
import keras
from keras.models import Sequential
from keras.layers import Dense, Dropout, Conv2D, MaxPooling2D, BatchNormalization, Flatten
#Import du modele VGG16 pre-entrainer sur le dataset inagenet
imagenet_model2 = tf.keras.applications.VGG16(weights = "imagenet",
include_top=False,
input_shape = (150, 150, 3),
pooling='max')
#Couches du modele VGG16
imagenet_model2.layers
# freeze feature layers and rebuild model
for l in imagenet_model2.layers:
l.trainable = False
#construction du model
model4 = [
tf.keras.layers.Flatten(),
tf.keras.layers.Dense(128, activation='relu'),
tf.keras.layers.Dense(12, activation='softmax')
]
#Jointure des deux modeles
model_using_pre_trained_one2 = tf.keras.Sequential( imagenet_model2.layers + model4 )
#choix de l'algo d'apprentissage
model_using_pre_trained_one2.compile(optimizer='adam',loss='sparse_categorical_crossentropy',metrics=['accuracy'])
#Apprentissage du modele
model_using_pre_trained_one2.fit(trainx2, trainy2, epochs=7,steps_per_epoch=len(trainx2))
#evaluation du modele
test_loss4, test_acc4 = model_using_pre_trained_one2.evaluate(testx2, testy2, verbose=2)
print('\nTest accuracy:', test_acc4)