Welcome!

Join our community of MMO enthusiasts and game developers! By registering, you'll gain access to discussions on the latest developments in MMO server files and collaborate with like-minded individuals. Join us today and unlock the potential of MMO server development!

Join Today!

Machine Learning Project - idea

Joined
Jun 8, 2007
Messages
1,985
Reaction score
490
I'm having a massive brain fart.

Usually I'm full of ideas, but I am requesting help.

Please give to me your original ideas for machine learning projects.
Disclaimer:
If I choose to work on your idea, you agree that the idea may be implemented on a GPL licensed project
published on github under s-p-n. You hold no rights to the idea, and are sharing for free with no liability (you won't go to jail for things I make) and no ownership (you won't take legal action against me for implementing the idea).


In fact, if you decide to give me your idea, I'll follow up with you and may ask further questions. I may attempt to get you involved as a partner. If I make any money, I may try and contact you to thank you, and again try to get you involved.

Further, anyone in the world is free to implement the idea you describe here. Completely open, completely libre.
 
Junior Spellweaver
Joined
Oct 27, 2008
Messages
165
Reaction score
89
You could try using the LBP descriptor to extract textures/patterns with a classifier.

There are a lot of descriptors that extract texture features from images, I chose LBP because it is the most simplest one, and can be used with any classifier.

The feature extracting is like this:
Input images => LBP Algorithm => Compute Histogram =>Pattern Matching =>Feature Vector

Before applying LBP Algorithm the input image can be pre-processed, to extract only image region of interest, you can also use multi-scale LBP to segment the image into different textures, or any other algorithm for segmentation.

If you use a SVM classifier it is recommended to use a Gaussian kernel, for the classifier, to get better results.

The training of the classifier is like this:
Training Set => Feature Extraction => Label Features => Train

And to use the classifier is like this:
Input image => Feature Extraction => Predict => Check Label

Another simple feature extraction algorithm is DCT

As applications:
- Feature Recognition(Facial, Hand gesture, etc)
 
Last edited:
Joined
Jun 8, 2007
Messages
1,985
Reaction score
490
Thanks Ghost! I'm still learning the terminology here. Basically, are you saying I can use LBP (or some other descriptor) to train features of interest- for example, hand gestures or facial expressions. Do you suppose Sign Language is a good candidate? I've always wanted to learn how to sign anyway, and the application space is large there with (for all I know) more demand than supply. I'm unsure about the limitations of machine learning. I'll be doing my training on my nvidia 950m using TensorFlow on ubuntu. Is that sufficient?

There's just so many things yet to be implemented.



Okay.

What if, we use Google Glasses to interpret sign language. The target user is a deaf individual. They could wear "smart glasses" that interpret their signs and speak for them. Also, we could use speech recognition to convert speech into sign language that is displayed on the user's glasses. Thus, 2-way communication for the hearing impaired.
 
Last edited:
Junior Spellweaver
Joined
Oct 27, 2008
Messages
165
Reaction score
89
For hand gestures is a good Idea, for sound, I'm not an expert but when having background noise(being in public), you will end in the cocktail party problem, where you have to separate the background sound.

As for what type of computer you need, any computer will do, if the image is too big you can resize the image to a smaller image for faster rough computing, then do the fine computing on a small region of the image.

For hand sign you will need to know edge detection to get the hand object boundary:
to do that you will need to extract the contours(it can be done by doing a simple threshold or canny edge detector).

As for hand contour feature you can extract: area, center, perimeter or the bounding rectangle of the hand.

Extracting shapes from an image:
1) Convert image to gray
2) Apply a canny filter on the image
- low pass filter(like gaussian filter) to remove the noise
- high pass filter(like sobel) to detect the edges
- threshold to make the edges more clear
3) Extract lines using the Hough transform
4) Build the shape(s) from the extracted lines

As a simple example of classification is the day and night classification:

Training pipeline:
Training images => Convert from RGB to HSV => Compute Average brightness => label average brightness as day or night =>Train Classifier

In the pipeline above image average brighness of the V channel will be the image feature for Day and Night, or you can use other than HSV color space like Lab, where L channel is the Lightness.

On prediction the pipeline is the same:
Image => Convert RGB to HSV => Compute Average brightness => Predict
 
Last edited:
Joined
Mar 22, 2009
Messages
1,232
Reaction score
502
A good idea that can be converted into a product, if there isn't already one that could sell well is to build a generic ML that can determine which influencers and which hashtags to use to promote certain products on social media and how effective a campaign is.
The data you'd get from the social media engines and possibly internal sales.
 
Back
Top