numbersrotationdl's Introduction

NumbersRotationDL

Implementation of ANN with number rotation

As a first step, the classification NN was implemented. I used CNN there.

try

anntest.m

As a second step, I build the regression network that detects an angle of digit rotation.

try

CNNregression.m

Then I implemented a script that freezed all the weights of the network.

FrozenANN.m

CustomLayer.m

This was a major problem for this project. I tried two different implementations. The fucntion:

failedRotation.m

doesn't even work. But I didn't delete it, hoping to fix.

The second function called

digitRotation.m

It performs bicubic interpolation. However, there is a need for clarification. Let's compare the behaviour of this function and MATLAB's imrotate()

However, here is rotation using my function:

We see that for some reason, the image was zoomed in by 10-15%. Nevertheless, the rotation happens.

customLayer.m

The customLayer was implemented in two ways. I tried to incorporate my function, but the problem with small numbers, so it needs to be optimized further. So, I eventually put imrotate() inside my custorLayer. Backprop is also implemented (in a trivial way).

Comments by the Reviewer

Good job, all goals accomplished
All networks performed as expected for the architecture chosen
Good knowledge of the layers and their proper placement in the model
I think you may have squeezed a bit more performance by allowing more epochs. Validation never terminated training, and I have a feeling that the net was not yet fully trained.
A very concise and readable implementation of rotation. It is much less complicated that the one ChatGPT wrote for us, and also that I found by Google search. As I recall, you used a book to study the algorithm. Would you please send me the reference?
NOTE: Used Emacs to indent your code, easier to read this way.