For example, search engine companies want to classify their image searches into 2 categories (which they already do that) such as: NSFW (nudity, porn, brutality) and safe to view pictures.
How can artificial neural networks achieve that, and at what success rate? Can they be easily mistaken?
The 2015 paper entitled "Applying deep learning to classify pornographic images and videos" applied various types of convnets for detecting pornography. The proposed architecture achieved 94.1% accuracy on the NPDI dataset, which contains 800 videos (400 porn, 200 non-porn "easy" and 200 non-porn "difficult"). More traditional computer vision methods achieved 90.9% accuracy. The proposed architecture also performs very well regarding the ROC curve.
There does not seem to exist any works regarding the other aspects of NSFW yet.
The problem detecting NSFW has been around for over two decades.
This study from 2005 about finding naked people, demonstrates a strategy for finding such images based on the color and texture properties to fetch an effective mask for skin regions attempting to group a human figure using geometric constraints on the human structure. This method demonstrated
60% precision and 52% recall on a test set of 138 uncontrolled images of naked people.
Here are a few figures from the study explaining the algorithm:
The following post contains visualizations of nudity for scientific purposes (hover to display):
A more recent approach is using convolutional networks. This study from 2014PDF demonstrated impressive classification performance based on the ImageNet dataset. It's not clear 'how and why they perform so well', however they can be used for classification of images with a very low error rate.
For further details, check: What convolutional neural networks look at when they see nudity.
You will find the code example and the heatmap for how convnets see NSFW in the above link.
