Algorithm to Identify Inappropriate Content on Photo Sharing Platform

Project Description

We are building a platform like Instagram where users will upload millions of images.  We need the ability to flag inappropirate images for moderator review.  From our platform perspective, inappropriate images are those that show nudity below the waist.  We are comfortable with the idea of receiving a small percentage of "false positives" since we will have moderators who can review them.

We see this as a two-phase project.  In the first phase, we would like an alogrithm that would classify these images into two buckets, appropriate and inappropriate, at a 90% confidence level. Down the road, we would have a second phase of development in which moderator intervention (signalling whether the algorithm was right or wrong in its classification) would train the alogrithm further based on whether the image was classified correctly.

Please be very specific about your approach.  Our short-term goal is to substantially reduce the number of images that our moderators have to review.  The algorithm must deliver the desired result for the project to be considered successful. We are unable to provide any training data since the site has not launched yet.