HyperNSFW-Caffe

Introduction

This is a Caffe implementation of Not Suitable for Work (NSFW) classification deep neural network. This model is now slightly more accurate than Yahoo's Open NSFW. As a bonus, this model executes ~4x faster. We can run classification with this model at 80msec on the mobile phone. For details, please read our blog.

Evaluate Models with a single image

Evaluate HyperNSFW:

For Python Code

python eval_image.py --proto model/mobilenet_v2_deploy.prototxt --model model/HyperNSFW.caffemodel --image ./test.jpg

Expected Outputs:

0.9302 - 'n16000001 sfw'
0.0698 - 'n16000002 nsfw'

For CPP Code

cd Prj-Cpp

Compile detection.cpp with OpenCV 3.4.1 or higher version.

You can also compile ncnn_mobilenet.cpp with ncnn framework.

Then test the code.

Finetuning on your own data

Prepare your own data and use related files in the train folder to Finetune on your own data.

Related Projects

HyperNSFW is trained with the following projects, we recommend you to take a look:

MobileNet-Caffe shicai/MobileNet-Caffe

识别APP

体验 Android APP：https://fir.im/HyperNSFW

Update (2018.12.03)

cpp code release.
Android Projects release.

bluelne/HyperNSFW