Introduction
Working with video datasets, notably with respect to detection of AI-based pretend 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 provided by OpenCV, magick, and keras.
Our method 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 may 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 photographs
Earlier than we go into an in depth clarification, readers ought to know that there isn’t any have to copy-paste code chunks. As a result of on the finish of the submit 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 lecturers.
It accommodates each actual and AI-generated pretend 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 numerous lengths. Our job is to find out the variety of pictures 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 information preparation.
Face detection
At first, face places should be decided by way of bounding bins, utilizing OpenCV. Then, magick is used to robotically extract them from all pictures.
# 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 = ''))
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 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(items = 256, activation = "relu") %>%
layer_dense(items = 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 submit exhibits the right way to 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 bins
- Construct a deep studying mannequin
Nevertheless, readers ought to know that the implementation of the next steps might drastically enhance mannequin efficiency:
- extract the entire 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 leads to the feedback part!
Thanks for studying!
Corrections
For those who see errors or need to counsel modifications, please create a difficulty on the supply repository.
Reuse
Textual content and figures are licensed below Inventive 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 below this license and could be acknowledged by a observe 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}
}