How can I increase my accuracy avoiding overfitting? CNN-Keras-VGG16

Question

As I asked in this question: Why are my predictions bad, if my accuracy in train is roughly 100% (Keras CNN) , my problem was Overfitting, so, I reduce the number of layers, and now I have this model:

vgg16_model = VGG16(weights="imagenet", include_top=True)

# (2) remove the top layer
base_model = Model(input=vgg16_model.input, 
                   output=vgg16_model.get_layer("block5_pool").output)

# (3) attach a new top layer
base_out = base_model.output
base_out = Reshape((25088,))(base_out)
# output layer: (None, 5)
top_preds = Dense(1, activation="sigmoid")(base_out)

# (4) freeze weights until the last but one convolution layer (block4_pool)
for layer in base_model.layers[0:14]:
    layer.trainable = False

# (5) create new hybrid model
model = Model(input=base_model.input, output=top_preds)

# (6) compile and train the model
sgd = SGD(lr=1e-4, momentum=0.9)
model.compile(optimizer=sgd, loss="binary_crossentropy", metrics=["accuracy"])

But, when I predict some images, the class 0 accuracy is roughly 96%, but the accuracy of class 1 is roughly 58%. So how can I increase the accuracy without overfitting.

I've trained my model with 700 images each class and to test I have 50 images each class.

Answer 1

Keras has support for image preprocessing. You can do lateral translations and shears on your images which will alter them but since the content is still the same your have more labeled data to train on. This is known as data augmentation.

Another thing you can do is introduce dropout in your network. This drops nodes in your network based on the probability you provide. This is done to prevent dependency of neurons on each other in your network.

Both of these where introduced in this paper. You can go through the reduce overfitting section to get a better idea and understanding.