Porn images detector with python, tensorflow, scikit-learn and opencv.
Two python porn images (nudity) detectors.
First one (pcr.py) use scikit-learn and opencv. I was able to get ~85% accuracy on markup with 1500 positive and 1500 negative samples. It use two machine-learned classifiers - one of them use HSV colors histogram, and another use SIFT descriptors.
Second one (nnpcr.py) uses tensorflow neural network. I was able to get ~90% accuracy on the same markup. It use 4 convolutional (3x3 filters) combined with max_pool (2x2) layers, one 1024 fully connected layer and a softmax classifier at the end.
This is my configuration, may be it can work with another library versions.
./pcr.py url http://example.com/img.jpg
from pcr import PCR
model = PCR()
model.loadModel('model.bin')
predictions = model.predict(['image1.jpg', 'image2.jpg', 'image3.jpg'])
print predictions
./nnpcr.py url http://example.com/img.jpg
from nnpcr import NNPCR
model = NNPCR()
model.loadModel('nnmodel.bin')
predictions = model.predict(['image1.jpg', 'image2.jpg', 'image3.jpg'])
print predictions
./pcr.py train
(to train opencv & sklearn) or ./nnpcr.py train
(for tensorflow one).After train finish you will see accuracy and you will get "model.bin" file with your trained model. Now you can use it to detect porn (see functions predictTest and predictUrl). I added a sample model (model.bin) - you can test it without training your own model, but I recomend you to gather some huge collection of images (eg, 50K) for best results.
Public domain (but it may use some patented algorithms, eg. SIFT - so you should check license of all used libraries).