Pornography/nudity Cascade training best size and parameters

asked 2017-11-27 23:26:45 -0500

ShayanFiroozi gravatar image

Hi everyone, first of all sorry fir my bad English.

i'm trying to train a classifier to detect porno/nude images.

i've collected 50,000 different porno images as positives and 100,000 images as negatives.

all i know is that a training in a scale like this could take 2,3 months !! so i have not any time for testing.

so i need the best parameters like -w -h -numPos - numNeg -numStages and so on...

also pictures i collected(positives and negatives) are extremely vary in resolution , from 320x200 to 1600x1200 !!!

should i resize them first ? if Yes , to what W and H of size ?

I appreciate any help , Thank you.

edit retag flag offensive close merge delete

Comments

1

forget that idea immediately. cascade classifiers are for rigid objects only

berak gravatar imageberak ( 2017-11-28 02:39:46 -0500 )edit
1

Maybe try a deep neural network like the VGG-16 or a more complex one (GoogLeNet), according to your training capacity...

kbarni gravatar imagekbarni ( 2017-11-28 11:35:01 -0500 )edit
2

or read this paper and so many papers...

LBerger gravatar imageLBerger ( 2017-11-28 11:37:24 -0500 )edit

you need a classification approach, not a detector ... you could train a boosted forest, but even then, OpenCV will not work. Go for available deep learning frameworks and transfer learning!

StevenPuttemans gravatar imageStevenPuttemans ( 2017-12-01 08:54:06 -0500 )edit