Introduction
Working with video datasets, notably with respect to detection of AI-based pretend objects, could be very difficult as a consequence of correct body choice and face detection. To strategy this problem from R, one could make use of capabilities provided by OpenCV, magick, and keras.
Our strategy consists of the next consequent steps:
- learn all of the movies
- seize and extract photographs from the movies
- detect faces from the extracted photographs
- 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:
However, magick is the open-source image-processing library that may assist to learn and extract helpful options from video datasets:
- Learn video recordsdata
- Extract photographs per second from the video
- Crop the faces from the pictures
Earlier than we go into an in depth clarification, readers ought to know that there is no such thing as a 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 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 out there.
The movies within the folders are within the format of mp4 and have numerous lengths. Our job is to find out the variety of photographs to seize per second of a video. We often 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?
Trying on the gif one can observe that some fakes are very straightforward to distinguish, however a small fraction seems fairly reasonable. That is one other problem throughout information preparation.
Face detection
At first, face areas have to be decided through bounding packing containers, utilizing OpenCV. Then, magick is used to mechanically extract them from all photographs.
# 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 crimson dashed line the field
imh = image_draw(image_read('frame_1.jpg'))
rect(rectX, rectY, x, y, border = "crimson",
lty = "dashed", lwd = 2)
dev.off()If face areas are discovered, then it is extremely 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 will shortly place all the pictures into folders and, utilizing picture mills, feed faces to a pre-trained Keras mannequin.
train_dir = 'fakes_reals'
width = 150L
top = 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,top),
batch_size = 10,
class_mode = "binary"
)
# Construct the mannequin ---------------------------------------------------------
conv_base <- application_vgg16(
weights = "imagenet",
include_top = FALSE,
input_shape = c(width, top, 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 reveals find out how to do video classification from R. The steps had been:
- Learn movies and extract photographs from the dataset
- Apply OpenCV to detect faces
- Extract faces through bounding packing 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 recordsdata
- 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 happy to attempt 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 wish to recommend adjustments, 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, except in any other case famous. The figures which have been reused from different sources do not fall underneath this license and might be acknowledged by a be aware 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/},
yr = {2020}
}
