如何优化预训练的 tensorflow CNN model 以减少过拟合并提高假人脸检测的整体准确度

Question

vgg19_net = VGG19(input_shape = (200,200,3), include_top = False, weights = 'imagenet')

model_vgg19 = Sequential()
model_vgg19.add(vgg19_net)

model_vgg19.add(Conv2D(128, 3, activation = 'relu', padding = 'same', kernel_initializer = 'he_uniform'))
model_vgg19.add(BatchNormalization())

model_vgg19.add(Conv2D(256, 3, activation = 'relu', padding = 'same', kernel_initializer = 'he_uniform'))
model_vgg19.add(BatchNormalization()) 

model_vgg19.add(Conv2D(512, 3, activation = 'relu', padding = 'same', kernel_initializer = 'he_uniform'))
model_vgg19.add(BatchNormalization())
model_vgg19.add(MaxPooling2D())

model_vgg19.add(Conv2D(256, 3, activation = 'relu', padding = 'same', kernel_initializer = 'he_uniform'))
model_vgg19.add(BatchNormalization())
model_vgg19.add(MaxPooling2D())

model_vgg19.add(Conv2D(128, 3, activation = 'relu', padding = 'same', kernel_initializer = 'he_uniform'))
model_vgg19.add(BatchNormalization()) 


model_vgg19.add(Flatten())
model_vgg19.add(Dense(128, activation = 'relu', kernel_initializer = 'he_uniform'))
model_vgg19.add(Dense(1, activation = 'sigmoid'))

model_vgg19.compile(loss='binary_crossentropy',optimizer=Adam(lr=1e-5), metrics=['accuracy'])
model_vgg19.summary()

Above is my python code wherein I am trying to add layers to a pre-trained VGG19 model. I have a dataset of real and fake faces and it's a classification question. I have normalized pixel values (between 0 and 1) for all images and I am using 80% as training data and 20% as validation data. I am currently getting an accuracy of 95-96% on validation but on my test set I am only getting around 91-92% accuracy. I am new to DL and still learning how to pre-process face images and build effective models. Please let me know if you see any abnormalities in my VGG19 implementation. I have used image augmentation as well but it hasn't proven to be useful. Please let me know how I can make this question better since I am new to SO. Also, the dataset is confidential so I cant share that. Sorry? Are there any pre-processing techniques I can use on my input images to better train my model?

Answer 1

I don't see any issues in your code. To improve this:- You Can Add Dropout To Reduce OverFitting , You Can Also Try To Use A ResNet Or A EfficientNet Version (Which Are Better At FakeFace Detection Than VGG); Most Of All, You Can Toy With The HyperParameters; Your LR is pretty low; Try Increasing It.. If U want advanced techniques, you can try Multiple Outputs Or Use Image Segmentation(It is highly unlikely to help increase your accuracy but multiple outputs might).

PS I recently worked on a similar dataset and managed to get 98% accuracy on the test set. ( I Used EffNetB2 )