Introduction
Working with video datasets, notably with respect to detection of AI-based faux objects, could be very difficult attributable to correct body choice and face detection. To strategy this problem from R, one could make use of capabilities supplied by OpenCV, magick, and keras.
Our strategy 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:
However, 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 photographs
Earlier than we go into an in depth clarification, 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.
Knowledge 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 incorporates each actual and AI-generated faux movies. The full measurement is over 470 GB. Nonetheless, the pattern 4 GB dataset is individually out there.
The movies within the folders are within the format of mp4 and have numerous lengths. Our process is to find out the variety of photos to seize per second of a video. We often 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?
Wanting on the gif one can observe that some fakes are very simple to distinguish, however a small fraction appears fairly life like. That is one other problem throughout knowledge preparation.
Face detection
At first, face areas must 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 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 = ''))
editedDeep studying mannequin
After dataset preparation, it’s time to construct a deep studying mannequin with Keras. We are able to rapidly 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 put up reveals learn 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
Nonetheless, readers ought to know that the implementation of the next steps could drastically enhance mannequin efficiency:
- extract all the 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 counsel modifications, please create a difficulty on the supply repository.
Reuse
Textual content and figures are licensed underneath Artistic Commons Attribution CC BY 4.0. Supply code is out there at https://github.com/henry090/Deepfake-from-R, until in any other case famous. The figures which were reused from different sources do not fall underneath this license and will 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/},
12 months = {2020}
}
