Introduction
Working with video datasets, significantly with respect to detection of AI-based pretend objects, may be very difficult as a consequence 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 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 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 pictures
Earlier than we go into an in depth rationalization, readers ought to know that there isn’t any must 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 varied teachers.
It comprises each actual and AI-generated pretend movies. The entire dimension is over 470 GB. Nonetheless, 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 photos to seize per second of a video. We often took 1-3 fps for each video.
Observe: 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 straightforward to distinguish, however a small fraction appears fairly life like. That is one other problem throughout information preparation.
Face detection
At first, face areas have to be decided by way of 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 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 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 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(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 photos from the dataset
- Apply OpenCV to detect faces
- Extract faces by way of bounding bins
- Construct a deep studying mannequin
Nonetheless, 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 at liberty to strive these choices on the Deepfake detection problem and share your ends in the feedback part!
Thanks for studying!
Corrections
When you see errors or need to recommend modifications, please create an issue 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 have been reused from different sources do not fall underneath this license and will 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,
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}
}
