[antpls]

Assuming you meant some SFW images are mistakenly blocked, it is an easier situation than with true negative.

Once an image tested positive with your first model, you could run a second more CPU intensive model (but also more accurate) on it.

[wongarsu]

Assuming you have a notably more accurate algorithm. It's a difficult problem even with infinite computing power.

[ndm000]

Great idea. MobileNetv2 is the model used currently due to it's small footprint (I used transfer learning on a dataset of NSFW and SFW images). A more robust model ran on all positives could improve performance overal.