Introduction

Detecting Not-Suitable-For-Work (NSFW) content is a high demand task in computer vision. While there are many types of NSFW content, here we focus on the pornographic images and videos.

The Yahoo Open-NSFW model originally developed with the Caffe framework has been a favourite choice, but the work is now discontinued and Caffe is also becoming less popular. Please see the description on the Yahoo project page for the context, definitions, and model training details.

This Open-NSFW 2 project provides a Keras implementation of the Yahoo model, with references to its previous third-party TensorFlow 1 implementation. Note that Keras 3 is compatible with TensorFlow, JAX, and PyTorch. However, currently this model is only guaranteed to work with TensorFlow and JAX.

A simple API is provided for making predictions on images and videos.

Installation

Tested with TensorFlow and JAX, for Python 3.9, 3.10, and 3.11.

A note on PyTorch:

The OpenNSFW 2 model can in fact be run on PyTorch, but the biggest issue is that the inference output on PyTorch is quite different from that on TensorFlow and JAX. The reason is still unknown. In addition, inference is much slower on PyTorch probably because of the issues discussed here, i.e., PyTorch uses channels_first for its image data format, but this model uses channels_last (as in TensorFLow and JAX), hence Keras has to convert the channel order back and forth at each layer. Therefore, at the moment it is not recommended to use PyTorch for this model.

The best way to install Open-NSFW 2 with its dependencies is from PyPI:

python3 -m pip install --upgrade opennsfw2

Alternatively, to obtain the latest version from this repository:

git clone git@github.com:bhky/opennsfw2.git
cd opennsfw2
python3 -m pip install .

Usage

Quick examples for getting started are given below. For more details, please refer to the API section.

Images

import opennsfw2 as n2

# To get the NSFW probability of a single image.
image_path = "path/to/your/image.jpg"

nsfw_probability = n2.predict_image(image_path)

# To get the NSFW probabilities of a list of images.
# This is better than looping with `predict_image` as the model will only be instantiated once
# and batching is used during inference.
image_paths = [
  "path/to/your/image1.jpg",
  "path/to/your/image2.jpg",
  # ...
]

nsfw_probabilities = n2.predict_images(image_paths)

Video

import opennsfw2 as n2

# The video can be in any format supported by OpenCV.
video_path = "path/to/your/video.mp4"

# Return two lists giving the elapsed time in seconds and the NSFW probability of each frame.
elapsed_seconds, nsfw_probabilities = n2.predict_video_frames(video_path)

Lower level with Keras

import numpy as np
import opennsfw2 as n2
from PIL import Image

# Load and preprocess image.
image_path = "path/to/your/image.jpg"
pil_image = Image.open(image_path)
image = n2.preprocess_image(pil_image, n2.Preprocessing.YAHOO)
# The preprocessed image is a NumPy array of shape (224, 224, 3).

# Create the model.
# By default, this call will search for the pre-trained weights file from path:
# $HOME/.opennsfw2/weights/open_nsfw_weights.h5
# If not exists, the file will be downloaded from this repository.
# The model is a `keras_core.Model` object.
model = n2.make_open_nsfw_model()

# Make predictions.
inputs = np.expand_dims(image, axis=0)  # Add batch axis (for single image).
predictions = model.predict(inputs)

# The shape of predictions is (num_images, 2).
# Each row gives [sfw_probability, nsfw_probability] of an input image, e.g.:
sfw_probability, nsfw_probability = predictions[0]

API

preprocess_image

Apply necessary preprocessing to the input image.

• Parameters:
  • pil_image (PIL.Image): Input as a Pillow image.
  • preprocessing (Preprocessing enum, default Preprocessing.YAHOO): See preprocessing details.
• Return:
  • NumPy array of shape (224, 224, 3).

Preprocessing

Enum class for preprocessing options.

• Preprocessing.YAHOO
• Preprocessing.SIMPLE

make_open_nsfw_model

Create an instance of the NSFW model, optionally with pre-trained weights from Yahoo.

• Parameters:
  • input_shape (Tuple[int, int, int], default (224, 224, 3)): Input shape of the model, this should not be changed.
  • weights_path (Optional[str], default $HOME/.opennsfw/weights/open_nsfw_weights.h5): Path to the weights in HDF5 format to be loaded by the model. The weights file will be downloaded if not exists. If None, no weights will be downloaded nor loaded to the model. Users can provide path if the default is not preferred. The environment variable OPENNSFW2_HOME can also be used to indicate where the .opennsfw2/ directory should be located.
  • name (str, default opennsfw2): Model name to be used for the Keras model object.
• Return:
  • tf.keras.Model object.

predict_image

End-to-end pipeline function from the input image to the predicted NSFW probability.

• Parameters:
  • image_path (str): Path to the input image file. The image format must be supported by Pillow.
  • preprocessing: Same as that in preprocess_image.
  • weights_path: Same as that in make_open_nsfw_model.
  • grad_cam_path (Optional[str], default None): If not None, e.g., cam.jpg, a Gradient-weighted Class Activation Mapping (Grad-CAM) overlay plot will be saved, which highlights the important region(s) of the (preprocessed) input image that lead to the prediction. Note that this feature is currently only supported by the TensorFlow backend.
  • alpha (float, default 0.8): Opacity of the Grad-CAM layer of the plot, only valid if grad_cam_path is not None.
• Return:
  • nsfw_probability (float): The predicted NSFW probability of the image.

predict_images

End-to-end pipeline function from the input images to the predicted NSFW probabilities.

• Parameters:
  • image_paths (Sequence[str]): List of paths to the input image files. The image format must be supported by Pillow.
  • batch_size (int, default 8): Batch size to be used for model inference. Choose a value that works the best with your device resources.
  • preprocessing: Same as that in preprocess_image.
  • weights_path: Same as that in make_open_nsfw_model.
  • grad_cam_paths (Optional[Sequence[str]], default None): If not None, the corresponding Grad-CAM plots for the input images will be saved. See the description in predict_image. Note that this feature is currently only supported by the TensorFlow backend.
  • alpha: Same as that in predict_image.
• Return:
  • nsfw_probabilities (List[float]): Predicted NSFW probabilities of the images.

Aggregation

Enum class for aggregation options in video frames prediction.

• Aggregation.MEAN
• Aggregation.MEDIAN
• Aggregation.MAX
• Aggregation.MIN

predict_video_frames

End-to-end pipeline function from the input video to predictions.

• Parameters:
  • video_path (str): Path to the input video source. The video format must be supported by OpenCV.
  • frame_interval (int, default 8): Prediction will be done on every this number of frames, starting from frame 1, i.e., if this is 8, then prediction will only be done on frame 1, 9, 17, etc.
  • aggregation_size (int, default 8): Number of frames for which their predicted NSFW probabilities will be aggregated. For instance, if a prediction will be done "on" frame 9 (decided by frame_interval), then it actually means prediction will be done on aggregation_size frames starting from frame 9, e.g., frames 9 to 16 if the size is 8. The predicted probabilities will be aggregated. After aggregation, each of these frames in that interval will be assumed the same aggregated probability.
  • aggregation (Aggregation enum, default Aggregation.MEAN): The aggregation method.
  • batch_size (int, default 8, upper-bounded by aggregation_size): Batch size to be used for model inference. Choose a value that works the best with your device resources.
  • output_video_path (Optional[str], default None): If not None, e.g., out.mp4, an output MP4 video with the same frame size and frame rate as the input video will be saved via OpenCV. The predicted NSFW probability is printed on the top-left corner of each frame. Be aware that the output file size could be much larger than the input file size. This output video is for reference only.
  • preprocessing: Same as that in preprocess_image.
  • weights_path: Same as that in make_open_nsfw_model.
  • progress_bar (bool, default True): Whether to show the progress bar.
• Return:
  • Tuple of List[float], each with length equals to the number of video frames.
    • elapsed_seconds: Video elapsed time in seconds at each frame.
    • nsfw_probabilities: NSFW probability of each frame. For any frame_interval > 1, all frames without a prediction will be assumed to have the NSFW probability of the previous predicted frame.

Preprocessing details

Options

This implementation provides the following preprocessing options.

• YAHOO: The default option which was used in the original Yahoo's Caffe and the later TensorFlow 1 implementations. The key steps are:
  • Resize the input Pillow image to (256, 256).
  • Store the image as JPEG in memory and reload it again to a NumPy image (this step is mysterious, but somehow it really makes a difference).
  • Crop the centre part of the NumPy image with size (224, 224).
  • Swap the colour channels to BGR.
  • Subtract the training dataset mean value of each channel: [104, 117, 123].
• SIMPLE: A simpler and probably more intuitive preprocessing option is also provided, but note that the model output probabilities will be different. The key steps are:
  • Resize the input Pillow image to (224, 224).
  • Convert to a NumPy image.
  • Swap the colour channels to BGR.
  • Subtract the training dataset mean value of each channel: [104, 117, 123].

Comparison

Using 521 private test images, the NSFW probabilities given by three different settings are compared:

• TensorFlow 1 implementation with YAHOO preprocessing.
• Keras implementation with YAHOO preprocessing.
• Keras implementation with SIMPLE preprocessing.

The following figure shows the result:

The current Keras implementation with YAHOO preprocessing can totally reproduce the well-tested TensorFlow 1 result, with small floating point errors only.

With SIMPLE preprocessing the result is different, where the model tends to give lower probabilities over the current test images.

Note that this comparison does not conclude which preprocessing method is "better", it only shows their discrepancies. However, users that prefer the original Yahoo result should go for the default YAHOO preprocessing.