An easiest model inspired from the Tensorflow tutorial of Deep Learning Specialisation of Prof. Andrew Ng.
The initial model had Layer 1 of 20 nodes, Layer 2 of 10 nodes followed by softmax layer of 5 nodes.
After tweaking the number of layers to 64-24-10, the acuracy given was about 0.997 and accuracy on test data was about 0.95
Next chosen number of nodes in layers changes to 256-32-10, which gave train accuracy 0.99995 and on test data gives 0.973
This shows how important the network structure is for getting good accuracy.
CNN Network
Next emplyoed model is CNN with network structure as follows.
CONV2D -> RELU -> MAXPOOL -> CONV2D -> RELU -> MAXPOOL -> FLATTEN -> FULLYCONNECTED
This network gives an train accuracy of 0.995571, slightly lesser than 256-32-10 network, but performs better than it on Test data. A better defined and studied network can give more better results.