Image Recognition (Composition by dust)

Author: dust
License: MIT
Date: 2010.10.24
Compatibility: 10.6
Categories:
Required plugins:
(none)

here is a basic image recognition kernel made in open cl. it is an unfiltered type of recognition meaning all RGBA values are checked pixel by pixel and the distance between training data and live input data is calculated. an average mean of the overall distance is calculated and normalized in a 0-1 range. zero being perfect.

you can record a short sample of you giving your self the peace sign or the middle finger its arbitrary then try and match. think sign language to speech or something like that as possible implementation.

you can try various thresh... settings for a looser image recognition. below .05 seems to be a decent all around setting. that means when the pixels are .05 apart in distance.

whats going on is i'm going pixel by pixel and calculating the distance between the training image and live image. if the distance is within the threshold then i keep the distance if not i make the distance 10 ultimately normalizing the average down to 0-1.

you need to record a training image. a 3 second sample seems to work good for me. also optimization's like inputing a vectorized image or something may help speed things up. as it is with the setting here i'm getting fps of n/a mostly.

feel free to record some samples or try different video feeds. this will work as gesture recognition as well. i will add the ability to save down various training sets so you so this could be used to for handwriting recognition or what ever.

PreviewAttachmentSize
recog.zip46.81 KB

dust's picture
Re: Image Recognition (Composition by dust)

don't really know how this is supposed to be done but it seems you can test any vectors distance not just a mesh's with open cl so that is pretty cool if you ask me.