Friday, September 12, 2014

POME: Research on neural network

Hi guys! It's been a long time. Well you know, procrastinating is awesome. LOL

The ANN

This week I managed to code the trainer program. Yay! What I have done is just made a GUI and integrated it with my already coded ANN. I have coded it last month.

Because I had the opportunity to play around with the ANN, then I tried to make a simple classifier network. Just to tested it out. Interestingly the ANN needed specific number to get it right. The more complex the network doesn't mean it's become more accurate.

For example, I gave 3 hidden layers with 50 hidden neurons each. When I tested it out the result was only 86% accuracy.
But when I used only 1 hidden layers with only 10 hidden neurons in it, the result of my classifier was 100% accuracy! Awesome!
These all still looks like a magic for me. Hahaha. But nevertheless it's still one same dataset, of course I expected exactly same result. For next experiment I'll try different person to write the character and tweak the parameters again.

Homography

Oh anyway, I've found out the problem from the last post of my data extractor (the strangely distorted and unaligned image at the corner). Looks like I need to use the homography.

Have you ever wonder how a person see through a camera but the camera itself produce a 2D image? Yep, it's all because homography plot the 3D world onto the 2D space. Can we reverse this effect? Of course we can! But how? Uh.. that's what I struggled on. I am still reading (and watching) about it. There's a good lecture about homography on YouTube!. I need this!!!


Humm.. That's it for now. I think now I can see my advisor again on this Monday. I asked by him to work on this part of research first. Well then, see ya!

Wednesday, September 3, 2014

Progress About My Paper 2 : The Corner

It was a lot of journey. I experimented to extract the letters from the form. But you know what? The problem is always lurking at the corner. Literally.

The Corner

I learned about the affine space from Tutsplus' tutorial . From the tutorial I managed to shear my image. The idea was like this. I have 3 points, which are top-left, bottom-left, and top-right. First, I set the top-left point as static point. Then I shear the bottom-left point so it's parallel to the top-left point. At last I shear the top-right point so it's parallel to the top-left point.

At the top-left-side part, the image looked correct.

But as the image going to the bottom-left corner the alignment began strange

From this result I concluded that I can't simply correct this with affine space. The image was on projective space.

I talked about this problem to my advisor. My advisor told me to ignore this problem for a moment. For the beginning step he told me to assume the form is perfectly aligned. I have to start to experimenting with Neural Network. Luckily I have a Wacom's pen tablet, so I can create handwritten form with it.

My extractor managed to export the XML format of letters. Here's how my extractor looks like.
To extract 52 characters it need about 5 seconds. Quite long eh? But it's enough for the moment. Now I need to create the program for ANN.

Wish me luck then. Ngoahahaha.

p.s: guhe lagi ngambil kursus bahasa Inggris. Lumayan nulis ginian, buat belajar structure.