UTILIZED EXIFTOOL TO AUTOMATICALLY EXTRACT METADATA FROM VIDEOS UPON UPLOADING, WHICH ASSISTS IN CATEGORIZING VIDEOS AND PINPOINTING LOCATIONS
THIS IS OUR DATA ArChiving Platform.
LEVERAGED MACHINE LEARNING TO AUTOMATICALLY BLUR THE FACES OF INDIVIDUALS IN IMAGES AND VIDEOS PRIOR TO ARCHIVING, EFFECTIVELY PROTECTING PRIVACY.
face blur technique
Gaussian Blur: Applies a fuzzy effect by smudging details with a kernel size that determines the smoothness level.
Pixelation: Converts the region into large, block-like pixels, mimicking a low-resolution image. Increasing pixel size merges more pixels into an averaged block, reducing detail and complicating original image reconstruction.
Random-sized Pixelation: Introduces variability by randomizing the pixelation grid size.
Pixel Scrambling: Randomizes pixel positions within the face region, thoroughly disrupting structural information
Superpixel Blurring:
Break the area into small groups called superpixels (formed based on both color and spatial proximity using the SLIC algorithm).
Make each group a single color based on the average color within that group. This simplifies the details and blurs the image.
The area still looks similar but the details are blurred out, making it harder to recognize specific features.
AUTOMATED FACE DETECTION
Process: Detects faces using machine learning, applies blurring techniques, and replaces detected face regions with their blurred counterparts.
Videos: Processes each video frame as an individual image, applies face blurring, and then reassembles the frames into a single video.
Model: YOLOv8 for Face Detection
Generates bounding boxes around each detected face.
Trained on the WIDER FACE dataset, which tests face detection algorithms under challenging conditions.
Performs fast and accurate face detection but struggles with faces that are really far away.
SECURITY
Since one of our main goals is to apply secure face blurring without compromising content integrity, superpixel blurring has proven to be the most effective method so far.
Moving forward, we plan to rigorously test the security of various face blurring techniques. This will include evaluating their resistance to sophisticated image reconstruction methods and assessing their performance in realistic scenarios to ensure they meet our privacy standards. We are also interested in exploring more advanced techniques for face blurring that can provide even better security while ensuring non-invasive blurring effects.
Overall, this is an ongoing project, and there is a vast landscape of techniques and methodologies we can explore, test, and develop further.
