Working with video datasets, significantly with respect to detection of AI-based faux objects, could be very difficult as a result of correct body choice and face detection. To strategy this problem from R, one could make use of capabilities provided by OpenCV, magick, and keras.

Our strategy consists of the next consequent steps:

  • learn all of the movies
  • seize and extract pictures from the movies
  • detect faces from the extracted pictures
  • crop the faces
  • construct a picture classification mannequin with Keras

Let’s rapidly introduce the non-deep-learning libraries we’re utilizing. OpenCV is a pc imaginative and prescient library that features:

Then again, magick is the open-source image-processing library that can assist to learn and extract helpful options from video datasets:

  • Learn video information
  • Extract pictures per second from the video
  • Crop the faces from the pictures

Earlier than we go into an in depth rationalization, readers ought to know that there is no such thing as a have to copy-paste code chunks. As a result of on the finish of the publish one can discover a hyperlink to Google Colab with GPU acceleration. This kernel permits everybody to run and reproduce the identical outcomes.

Knowledge exploration

The dataset that we’re going to analyze is supplied by AWS, Fb, Microsoft, the Partnership on AI’s Media Integrity Steering Committee, and varied lecturers.

It comprises each actual and AI-generated faux movies. The whole measurement is over 470 GB. Nevertheless, the pattern 4 GB dataset is individually accessible.

The movies within the folders are within the format of mp4 and have varied lengths. Our job is to find out the variety of pictures to seize per second of a video. We often took 1-3 fps for each video.

Be aware: Set fps to NULL if you wish to extract all frames.

video = magick::image_read_video("aagfhgtpmv.mp4",fps = 2)
vid_1 = video[[1]]
vid_1 = magick::image_read(vid_1) %>% image_resize('1000x1000')

We noticed simply the primary body. What about the remainder of them?

Trying on the gif one can observe that some fakes are very simple to distinguish, however a small fraction appears to be like fairly life like. That is one other problem throughout information preparation.

Face detection

At first, face areas should be decided by way of bounding containers, utilizing OpenCV. Then, magick is used to mechanically extract them from all pictures.

# get face location and calculate bounding field
unconf <- ocv_read('frame_1.jpg')
faces <- ocv_face(unconf)
facemask <- ocv_facemask(unconf)
df = attr(facemask, 'faces')
rectX = (df$x - df$radius) 
rectY = (df$y - df$radius)
x = (df$x + df$radius) 
y = (df$y + df$radius)

# draw with crimson dashed line the field
imh  = image_draw(image_read('frame_1.jpg'))
rect(rectX, rectY, x, y, border = "crimson", 
     lty = "dashed", lwd = 2)

If face areas are discovered, then it is rather simple to extract all of them.

edited = image_crop(imh, "49x49+66+34")
edited = image_crop(imh, paste(x-rectX+1,'x',x-rectX+1,'+',rectX, '+',rectY,sep = ''))

Deep studying mannequin

After dataset preparation, it’s time to construct a deep studying mannequin with Keras. We will rapidly place all the pictures into folders and, utilizing picture mills, feed faces to a pre-trained Keras mannequin.

train_dir = 'fakes_reals'
width = 150L
peak = 150L
epochs = 10

train_datagen = image_data_generator(
  rescale = 1/255,
  rotation_range = 40,
  width_shift_range = 0.2,
  height_shift_range = 0.2,
  shear_range = 0.2,
  zoom_range = 0.2,
  horizontal_flip = TRUE,
  fill_mode = "nearest",

train_generator <- flow_images_from_directory(
  target_size = c(width,peak), 
  batch_size = 10,
  class_mode = "binary"

# Construct the mannequin ---------------------------------------------------------

conv_base <- application_vgg16(
  weights = "imagenet",
  include_top = FALSE,
  input_shape = c(width, peak, 3)

mannequin <- keras_model_sequential() %>% 
  conv_base %>% 
  layer_flatten() %>% 
  layer_dense(models = 256, activation = "relu") %>% 
  layer_dense(models = 1, activation = "sigmoid")

mannequin %>% compile(
  loss = "binary_crossentropy",
  optimizer = optimizer_rmsprop(lr = 2e-5),
  metrics = c("accuracy")

historical past <- mannequin %>% fit_generator(
  steps_per_epoch = ceiling(train_generator$samples/train_generator$batch_size),
  epochs = 10

Reproduce in a Notebook


This publish reveals how you can do video classification from R. The steps had been:

  • Learn movies and extract pictures from the dataset
  • Apply OpenCV to detect faces
  • Extract faces by way of bounding containers
  • Construct a deep studying mannequin

Nevertheless, readers ought to know that the implementation of the next steps could drastically enhance mannequin efficiency:

  • extract the entire frames from the video information
  • load completely different pre-trained weights, or use completely different pre-trained fashions
  • use one other know-how to detect faces – e.g., “MTCNN face detector”

Be happy to attempt these choices on the Deepfake detection problem and share your ends in the feedback part!

Thanks for studying!


When you see errors or wish to counsel modifications, please create an issue on the supply repository.


Textual content and figures are licensed below Artistic Commons Attribution CC BY 4.0. Supply code is out there at, until in any other case famous. The figures which were reused from different sources do not fall below this license and might be acknowledged by a word of their caption: “Determine from …”.


For attribution, please cite this work as

Abdullayev (2020, Aug. 18). Posit AI Weblog: Deepfake detection problem from R. Retrieved from

BibTeX quotation

  writer = {Abdullayev, Turgut},
  title = {Posit AI Weblog: Deepfake detection problem from R},
  url = {},
  12 months = {2020}

Leave a Reply

Your email address will not be published. Required fields are marked *