2

I am doing my final year project in android. The user can write the alphabet and the application will check whether it's written correctly or not.

i am planning to use neural network. The problem is that for a letter say "A", the attached images shown below will be considered as alphabet "A" even it is not "A". I trained the letter "A" with correct image but it will consider all the images as "A", I believe that all the neural network methods will check for the occurrence anywhere in the image.

Since my application is not a handwriting recognition tool but rather an evaluation tool for teachers: which algorithm will meet my requirements?

What I should need is an algorithm which will execute before the neural network algorithm to check any distortion, elongation or overwriting of character there is.

Can i use freeman chain code method? Is it faster than neural network in the android platform?

enter image description here

divanov
  • 6,173
  • 3
  • 32
  • 51
Shanij P.S
  • 164
  • 3
  • 12
  • 1
    When doing some image processing and recognition, first you need to determine variation of your input. That means that you need to know what kind of image sizes, contrast, brightness etc. Once you established your input, for neural networks, you first create training series, like 100 different images, or whatever is your need, and then you can train it for recognition. It seems to me like you got your problem and someone just told you, "Yeah you can use neural networks for that", yes you can but you need to determine how big is your network and how are you going to train it. – Marko Lazić Jan 19 '14 at 14:07
  • can u propose an algorithm to detect deviations from normal characters...like counting pixels, overlaping etc...but i think it will not work – Shanij P.S Jan 19 '14 at 14:35
  • You can train your network using wrong examples as well and teach her not to recognise those as As. Maybe you can find a threshold under which the image is not a valid character. For example, I found that my network has an average confidency of about 90% when it correctly classifies an image as digit and of 40% when the classification is wrong – BlackBear Jan 19 '14 at 14:41
  • yes i know this methods.. but my application needs English, Malayalam , Tamil and Kannada letters so it is not possible to train for all the possibilities. first i have to check geometric property if it passes the give the input to neural network – Shanij P.S Jan 19 '14 at 14:48

0 Answers0