In this repository you can find an effective CNN model for detecting NSFW content in images. It achieves a 97.70% accuracy on the pornography-2k dataset. Apart from the provided code you need to reproduce the model, you can also use this model as python module (see below).
First install the python module:
pip install --upgrade nsfwdetection
Then use the model by:
#import the model
from nsfw_detector.model import Model
# initialize the model
net = Model()
# make a prediction
output = net.predict(<imagepath>)
# make multiple predictions
output = net.predict([<imagepath>, <imagepath>])
The model's output is as follows: {'image_path': {'Label': SFW or NSFW , 'Score': 0 to 1 }}
Using the provided code you can retrain the model. Regarding the training/testing data, the corresponding splits are provided in folder "splits". You can find and download the necessary data from here (this is a private dataset, due to its content and you should send a request to gain access - find more info in the link) and here. After downloading the 2 datasets, you should run the
python script1_extract_image_frames.py
to extract the frames from the videos of the pornography-2k dataset and then you can run the
bash train_exps.sh
to start training the model.
As you can see in the following attention maps, the model exhibits high accuracy in terms of the image regions it focuses. (Censored images)
