How can I increase the accuracy of my image classification keras model in Python?

Question

I am trying to classify a group of bee images into two classes - bumble bee and honey bee with the resulting format being a CSV file like -

id,bumble_bee,honey_bee

20000,0.75, 0.25.

I have a running model, but the accuracy is very low and I've tried a bunch of different things such as adding a base_model like VGG16 or InceptionV3, adjusting the epochs, adjusting the optimizer type... and I just haven't noticed much difference. All of my changes still result in an accuracy of around 70-79%.

How can I increase the accuracy of my model?

Here is my code:

from keras.preprocessing.image import ImageDataGenerator
from keras.models import Sequential
from keras.layers import Conv2D, MaxPooling2D
from keras.layers import Activation, Dropout, Flatten, Dense
from keras import backend as K

# dimensions of our images.
img_width, img_height = 200, 200

train_data_dir = 'data/train/'
validation_data_dir = 'data/validation/'
nb_train_samples = 2978
nb_validation_samples = 991
epochs = 50
batch_size = 25

if K.image_data_format() == 'channels_first':
    input_shape = (3, img_width, img_height)
else:
    input_shape = (img_width, img_height, 3)

model = Sequential()
model.add(Conv2D(32, (3, 3), input_shape=input_shape))
model.add(Activation('relu'))
model.add(MaxPooling2D(pool_size=(2, 2)))

model.add(Conv2D(32, (3, 3)))
model.add(Activation('relu'))
model.add(MaxPooling2D(pool_size=(2, 2)))

model.add(Conv2D(64, (3, 3)))
model.add(Activation('relu'))
model.add(MaxPooling2D(pool_size=(2, 2)))

model.add(Flatten())
model.add(Dense(64))
model.add(Activation('relu'))
model.add(Dropout(0.5))
model.add(Dense(2))
model.add(Activation('softmax'))

model.compile(loss='categorical_crossentropy',
              optimizer=optimizers.SGD(lr=1e-4, momentum=0.9),
              metrics=['accuracy'])

# this is the augmentation configuration we will use for training
train_datagen = ImageDataGenerator(
    rescale=1. / 255,
    shear_range=0.2,
    zoom_range=0.2,
    horizontal_flip=True)

# this is the augmentation configuration we will use for testing:
# only rescaling
test_datagen = ImageDataGenerator(rescale=1. / 255)

train_generator = train_datagen.flow_from_directory(
    train_data_dir,
    target_size=(img_width, img_height),
    batch_size=batch_size,
    class_mode='categorical')

validation_generator = test_datagen.flow_from_directory(
    validation_data_dir,
    target_size=(img_width, img_height),
    batch_size=batch_size,
    class_mode='categorical')

model.fit_generator(
    train_generator,
    steps_per_epoch=nb_train_samples // batch_size,
    epochs=epochs,
    validation_data=validation_generator,
    validation_steps=nb_validation_samples // batch_size)

model.save_weights('thirdtry.h5')

pred = model.predict(pred_images)
np.savetxt('resultsfrom3no.csv',pred)

Here is an example of its output:

Found 2978 images belonging to 2 classes. Found 991 images belonging to 2 classes.

Epoch 1/50 119/119 [==============================] - 238s 2s/step - loss: 0.5570 - acc: 0.7697 - val_loss: 0.5275 - val_acc: 0.7908

Epoch 2/50 119/119 [==============================] - 237s 2s/step - loss: 0.5337 - acc: 0.7894 - val_loss: 0.5270 - val_acc: 0.7908

Epoch 3/50 119/119 [==============================] - 282s 2s/step - loss: 0.5299 - acc: 0.7939 - val_loss: 0.5215 - val_acc: 0.7908

Have you considered the possibility that the classification task is actually too difficult? This particular difference is hard for humans to notice, as far as I know. Is there a way to get more data? Is anything misclassified (noise)? — Henry
Does training accuracy also stagnate after a few epochs? Show the training/validation results. Do you only have 119 samples to train on? — KonstantinosKokos
@KonstantinosKokos - Accuracy, validation accuracy, validation loss, and loss generally stagnate the entire time. The 119 is because there are 2978 training samples and a batch size of 25 (2978/25 = 119.12). — kaecvtionr
@Henry - The generators are suppose to assist in creating more data. — kaecvtionr
@kaecvtionr I see the image generators, but the possibility still stands. I realize I wasn't terribly helpful though :) — Henry

pronot pronot · Accepted Answer · 2018-06-16T10:10:28

Use an image-net pre-trained VGGNet for feature extraction and build your classifier on top of it with Dense layers. You should set the image size (w x h) to original network input.

Since you have only two classes, it will be better to use binary class mode for data generators, binary cross entropy as loss function and sigmoid as final activation function.

First try without data augmentation for a quick test and then use augmentation to see if it helps.

While training your classifier network, start with a respectively high learning rate (such as 0.001) and experiment with lower rates (1e-4, 1e-5, etc...) if the acc value gets sticked.

Also in your code, decreasing the learning rate and playing with the momentum value will probably help.

How can I increase the accuracy of my image classification keras model in Python?

1 Answers