NSFW IMAGE CLASSIFICATION USING EFFICIENTNET

We are in the century of the internet with the amount of data used each year reaching to 40+ zettabytes [1]. According to some research groups, over 30% of the internet users have experienced unwanted exposure to NSFW content through banner ads, pop-up ads, misdirected links or emails. Most of the content today is publicly viewable, and can be accessed from anywhere, be it the workplace, home or even school computers. But, not all of the content is appropriate for all users, especially children. An example such content is pornographic images which should be restricted to adults. Besides, these images are Not Safe For. Recently, many new convolutional neural network architectures have been successfully applied to many computer vision problems with efficient training and faster inference times in mind. One such architecture is EfficientNet. This work proposes to classify NSFW images using EfficientNetV2. A dataset of images is built by scraping Reddit, Imgur and Pornpics. Pretrained EfficientNetB5 model from Keras is used to first determine the best image size and then a final model is trained on 450×450 images with the final accuracy being 93%.