Introduction

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

Our method consists of the next consequent steps:

• learn all of the movies
• seize and extract photos from the movies
• detect faces from the extracted photos
• 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:

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

• Learn video information
• Extract photos 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 isn’t a have to copy-paste code chunks. As a result of on the finish of the put up one can discover a hyperlink to Google Colab with GPU acceleration. This kernel permits everybody to run and reproduce the identical outcomes.

Information exploration

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

It comprises each actual and AI-generated faux movies. The entire 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 numerous lengths. Our activity is to find out the variety of photos to seize per second of a video. We normally 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?

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

Face detection

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

# get face location and calculate bounding field
library(opencv)
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 pink dashed line the field
imh  = image_draw(image_read('frame_1.jpg'))
rect(rectX, rectY, x, y, border = "pink", 
     lty = "dashed", lwd = 2)
dev.off()

If face areas are discovered, then it is vitally 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 = ''))
edited

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 turbines, 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",
  validation_split=0.2
)


train_generator <- flow_images_from_directory(
  train_dir,                  
  train_datagen,             
  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(
  train_generator,
  steps_per_epoch = ceiling(train_generator$samples/train_generator$batch_size),
  epochs = 10
)

Conclusion

This put up exhibits find out how to do video classification from R. The steps have been:

• Learn movies and extract photos 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 might drastically enhance mannequin efficiency:

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

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

Thanks for studying!

Corrections

In case you see errors or need to recommend adjustments, please create a problem on the supply repository.

Reuse

Textual content and figures are licensed beneath Inventive Commons Attribution CC BY 4.0. Supply code is offered at https://github.com/henry090/Deepfake-from-R, except in any other case famous. The figures which were reused from different sources do not fall beneath this license and may be acknowledged by a notice of their caption: “Determine from …”.

Quotation

For attribution, please cite this work as

Abdullayev (2020, Aug. 18). Posit AI Weblog: Deepfake detection problem from R. Retrieved from https://blogs.rstudio.com/tensorflow/posts/2020-08-18-deepfake/

BibTeX quotation

@misc{abdullayev2020deepfake,
  creator = {Abdullayev, Turgut},
  title = {Posit AI Weblog: Deepfake detection problem from R},
  url = {https://blogs.rstudio.com/tensorflow/posts/2020-08-18-deepfake/},
  12 months = {2020}
}