Image and content analysis to determine the differences between a tasteful photograph of a person, a swimsuit photograph, a nude photograph, depictions of pornography... as far as I know is nowhere near sophisticated enough to do in software alone.
Fortunately crowdsourcing should be useful here, as @ammoQ suggested in a comment. However I don't believe members of 4chan or any other forum would appreciate the vast number of non-pornographic images, such as generic web graphics for buttons, frames, advertisements, etc. being posted.
My recommendation would be to look into existing crowdsourcing solutions, such as Amazon Mechanical Turk. (However the terms of service may explicitly prohibit the involvement of pornographic content, so be advised you might have to find another solution or roll your own.)
To make crowdsourcing feasible, your software should be prepared to do some or all of the following:
- Store information that links the content with the computer it came from
- Identify exact duplicates across the entire inventory and remove them (but origin information is retained)
- Downsample images to some dimension, perhaps 320x200, which is sufficient to identify the content of the image without retaining unnecessary detail and wasting storage space/bandwidth
- Create still images of video content at some regular interval and apply the same downsampling rule
Finally, the database of reduced images that represent the original image and video content is checked by users (or a designated team if you have the resources) according to your company's code of conduct. The program or interface might show a single image at a time, or a screen of thumbnails--whatever you deem best to obtain accurate information.
The identity of the computer from which images came should absolutely be secret and unknown to the persons evaluating the data. Additionally it should be randomized and each image probably checked more than once to remove bias.
The same technique could be used for text, but first the content could be scored by keyword rankings which remove the bulk of text from crowdsource review. Classifying a long document will of course be more time consuming than classifying an image.