I tried to implement a three layer CNN with Batch Normalisation and MaxPooling. I got an Accuracy of 90.33 % after 5000 iterations. I added Batch normalisation after every later to accelerate the performance.

The Architecture is as follows- Convolutional Layer 1 with output feature map 16 and 5*5 kernel ReLU Activation MaxPooling 1 BatchNormalisation 1

Convolutional Layer 2 with output feature map 32 and 5*5 kernel ReLU Activation MaxPooling 2 BatchNormalisation 2

Convolutional Layer 3 with output feature map 64 and 5*5 kernel ReLU Activation MaxPooling 3 BatchNormalisation 3

Fully Connected Layer.