Menu

Adult Image Recognizer using HTM

softnhard
2013-03-28
2013-05-19
  • softnhard

    softnhard - 2013-03-28

    I've developed an Adult image recognizer based on NuPIC 1.7 and it works fin achieving up to 87% in precision and 83% in recall. It is in use in my company in order to categorize images. I'm preparing a paper related explaining about network and parameters used.

     
    • David Ragazzi

      David Ragazzi - 2013-03-28

      Hi softnhard,

      Thanks for your contribution, but I wonder why you still is developing in NuPIC since it is obsolete as own Hawkins said. NuPIC don't allow hierarchy and (to me) is much more complex for developing than CLA. Due the vision have biologically more than one region, CLA and its hierarchy is the most recommended to create applications to handle images.

      David

       
  • softnhard

    softnhard - 2013-03-28

    This project has been started in 2011 At that time there was not any good implementation or explanation of the CLA other than the white paper published in Dec 2010. So after 6 months of working on primary CLA implementations we switched to NuPIC and could train and use it effectively. In NuPIC 1.7 there is a toolkit named VisionDemo which has consisted base of our work. Our implementation is consisted of a 6+2 layer where first layer is Gabor Sensor next 6 layer are alternating Spatial and Temporal poolers each pair processing input from previous pair and last layer is an SVM classifier. I'll try in (I hope near) future implement our software on the new CLA based OpenHTM code. But currently results are surprising and our Adult Image Recognizer is competing with other enterprise solutions showing extraordinary with greater accuracy and performance.

     

    Last edit: softnhard 2013-03-28
    • David Ragazzi

      David Ragazzi - 2013-03-29

      Ok.. It's nice know you want use CLA, we're improving the algorithm and soon you will can use it as core.

       

      Last edit: David Ragazzi 2013-03-29
  • Magnus Wootton

    Magnus Wootton - 2013-05-19

    When recognizing images, you have to OR together all the perspectives and positions. (all the different symantics) and it should work! :D

    feeding back from the position inside the OR, you should then be able to locate the position of the thing in the vision!

     

    Last edit: Magnus Wootton 2013-05-19

Log in to post a comment.