Introduction
Working with video datasets, notably with respect to detection of AI-based faux 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 rapidly introduce the non-deep-learning libraries we’re utilizing. OpenCV is a pc imaginative and prescient library that features:
Alternatively, magick is the open-source image-processing library that can assist to learn and extract helpful options from video datasets:
- Learn video information
- 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 must 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.
Knowledge exploration
The dataset that we’re going to analyze is supplied by AWS, Fb, Microsoft, the Partnership on AI’s Media Integrity Steering Committee, and varied teachers.
It accommodates each actual and AI-generated faux movies. The entire 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 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?
Wanting on the gif one can observe that some fakes are very simple to distinguish, however a small fraction seems fairly reasonable. That is one other problem throughout knowledge preparation.
Face detection
At first, face areas should be decided by way of bounding bins, 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 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 areas are discovered, then it is vitally 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 pictures 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(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 publish reveals how one can do video classification from R. The steps have been:
- Learn movies and extract photographs 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 attempt these choices on the Deepfake detection problem and share your ends in the feedback part!
Thanks for studying!
Corrections
If you happen to see errors or need to counsel adjustments, please create a difficulty on the supply repository.
Reuse
Textual content and figures are licensed beneath Inventive Commons Attribution CC BY 4.0. Supply code is obtainable 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 beneath 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}
}