This AI can reconstruct motion blurred Human faces

We all know how irritating it can be when you are taking pictures of some high-speed bike ride or hockey match with your cellphones and don’t look at the focus feature of your camera which usually results in a gallery full of unusable blurry images.

We all have been through these problems, and to ease people facing such situations scientists at the Inception Institute of Artificial Intelligence in the United Arab Emirates, the Beijing Institute of Technology and Stony Brook University all worked together to develop an Artificial Intelligence (AI) system that could help the people to remove the blur from pictures during post-production.

According to the paper from the researchers, this new AI-supported system is also aware of back ground and foreground and can easily deblur human faces as well along with that, the system also works pretty much against the state-of-the-art motion deblurring methods as well.

Detailed insight on this new AI system

When a person tries to capture a picture of some high-speed object, due to motion between the camera and the objects the results end up with different types of degradation in the foreground and background and due to a space between the image planes it usually ends up with varied motions in subjects as well.

This is where the new AI-based model come in handy as it learns about human and background masks and later on leverages the masks to detect the foreground and background area. To train this model, the researchers compiled a data set of blurry images with ground truth shape images which included thousands of outdoor scenes and images of complex backgrounds and foreground motions and sizes.

This model was fed with Human Aware Image Deblurring (HIDE) pairs and each of those pairs was trained via human detection models which resulted in some rough accurate around subjects which were later refined by human annotators.


This model by the researchers uses the Nvidia Titan X graphics card and the model is trained with both HIDE and GoPro Hero data set of video frames with more than 10,742 images.

According to the researchers, their AI achieved state-of-the-art performance with its dynamic deblurring results as compared to other baselines. The deblurring feature enabled the researchers to reconstruct the images with explicit structure and semantic details.

According to the paper from the researchers, this design enables a more unified and human-aware deblurring network that explicitly separates the human-related images and background blur.

The method used by this machine-learning model enables the users to detect images with a variety of motion patterns involving humans as well with better results for both foreground and backgrounds.

Previous works are done by researchers

This isn’t the first time that the researchers worked on a method to clean messy images through AI but instead Nvidia, MIT and Aalto University also proposed various machine learning techniques to reduce image noise and a model from Chinese smartphone manufacturer Xiaomi was also launched that helps its users to restore details of images and to enhance the colors of pictures with extreme exposures.



Read next: Meet Meena, Google's most advanced conversational Artificial intelligence

No comments:

Post a Comment