...

1. Software Environment
2. Data Generation
3. Codewords Generation
   1. CPU
   2. GPU
      
4. Support Vector Machine
5. Convolutional Neural Network (pending)
6. My Understanding 

1.Software Environment

The required software are listed below:

...

http://docs.python-guide.org/en/latest/dev/virtualenvs/

2.Data Generation

6.My Understanding

I think this problem is not one of those that a typical neural network may be good at. 

I believe that currently, neural network is at most as good as human. Or it may be better than human sometimes but that is only because it may use more resource and be more stable. So if a task is in principle impossible for human, then it is highly poossible that no neural network can finish it. It seems to be that this is one of that kind of task.

Some of the images from the vcc are in principle impossible for human to recognize when then noise is too high.  We are able to pinpoint the spot only after the background subtraction. Thus the correlation between different images is crucial for this task.

Currently we don't have many images available for training. Thus I use two ways to generate new images.

• For images with a beam, first shift the beam spot to the center of the image, then rotate the image around the center, then cut out a small patch of images that contains the beam. Then arbitrarily chose a position paste the small patch to the new position. Cut out the small patch of image that is covered by the beam spot, and paste it to the original position of the beam spot. This proccedure aims to minimize the influence of our manipulation. It should be notice that to do this, one have to choose a large enough patch. Otherwise, it may not be able to cover the original beam spot.
• For all images, one can merely shift the whole images one or two pixels along x or y axis. In this way, one can create 24 times more data.

3.Codewords Generation  

Of course, the correlation is also important in classification problem. But the situaltion is different here. We need an exact correlation to do the right background subtraction.