Artificial intelligence is turning the 鈥淓nhance!鈥� trope on TV crime dramas 鈥� where a blurry image of a suspect turns into a clear photograph 鈥� into reality.
Vishal Patel at John Hopkins University in the US and his colleagues have developed an AI that can automatically deblur photographs of people鈥檚 faces.
Advertisement
The technology could eventually be used to improve facial recognition on long-distance surveillance images, such as those taken from a drone, says Patel.
The AI was developed in collaboration with a researcher at Adobe, a firm that makes photo editing software, and could also be used as a tool to automatically fix blur in camera photos.
Deblurring photos is challenging because many different types of blur exist, and it is difficult to determine which kinds are present in an image. Examples include motion blur, which occurs when a photo is taken while a camera is moving, or vice versa 鈥� when a person is moving but the camera is stationary.
The team trained the AI by providing it with thousands of photos of people鈥檚 faces, both clear and blurry. In each picture, different facial features were labelled, such as the eyes, nose and mouth, which previous AIs have had difficulty deblurring.
The AI was trained to deblur individual features including skin and hair, and then combine the deblurred features into a final clear photograph.
The team tested the AI鈥檚 ability to deblur 16,000 images that had been previously blurred and compared the algorithm鈥檚 result to the original clear image.
Its performance was assessed using a structural similarity index, where a score of 1 indicates exact similarity between two images. The algorithm scored 0.96.
Blur is just one of several factors that can affect the quality of an image, particularly if it is being used for forensic purposes.
These include different weather and light conditions, or a person being partially obscured by objects or other people, says Shufan Yang at the University of Glasgow. High-speed winds and magnetic fields can also interfere with cameras, she says.
鈥淭he other thing that you see in CCTV cameras is low resolution,鈥� says Patel. As a result, the images are often small in size and lack detail. To counter this problem, neural networks that create image details have been developed to .
Many ethical questions need to be answered before such technologies can be used forensically, given the risk that an AI could generate a clear picture of a suspect that is different to what the person actually looks like.
Combined with inaccurate or racially biased facial recognition algorithms, this could lead to the wrong person being identified.
Reference:听
