Introduction
Working with video datasets, notably with respect to detection of AI-based faux objects, may be very difficult as a result of 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 shortly 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 photos per second from the video
- Crop the faces from the photographs
Earlier than we go into an in depth rationalization, readers ought to know that there isn’t any 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.
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 varied lecturers.
It comprises each actual and AI-generated faux movies. The overall measurement is over 470 GB. Nevertheless, the pattern 4 GB dataset is individually obtainable.
The movies within the folders are within the format of mp4 and have varied 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.
Notice: 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 straightforward to distinguish, however a small fraction seems to be fairly practical. That is one other problem throughout knowledge preparation.
Face detection
At first, face places must be decided through bounding bins, utilizing OpenCV. Then, magick is used to robotically 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 purple dashed line the field
imh = image_draw(image_read('frame_1.jpg'))
rect(rectX, rectY, x, y, border = "purple",
lty = "dashed", lwd = 2)
dev.off()
If face places are discovered, then it is vitally straightforward 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 shortly place all the photographs 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",
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 publish reveals methods to do video classification from R. The steps had been:
- Learn movies and extract photos from the dataset
- Apply OpenCV to detect faces
- Extract faces through bounding bins
- 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 expertise to detect faces – e.g., “MTCNN face detector”
Be at liberty to strive these choices on the Deepfake detection problem and share your leads to the feedback part!
Thanks for studying!
Corrections
Should you see errors or need to recommend adjustments, please create a problem on the supply repository.
Reuse
Textual content and figures are licensed underneath Artistic Commons Attribution CC BY 4.0. Supply code is accessible at https://github.com/henry090/Deepfake-from-R, except in any other case famous. The figures which have been reused from different sources do not fall underneath 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, writer = {Abdullayev, Turgut}, title = {Posit AI Weblog: Deepfake detection problem from R}, url = {https://blogs.rstudio.com/tensorflow/posts/2020-08-18-deepfake/}, yr = {2020} }