Here I have done a little experiment. I've created two jupyter notebooks to compare the performance of a model with no hidden layer (equivalent to a binomial logistic regression), the second model has home hidden layers, the third is bigger than the previous one and has Dropout layers to avoid overfitting and VGG16 pretrained model.
Check the notebooks to compare the performance of the models.