Retina-Inspired Semiconductor Sensor Boosts AI Recognition of Blurry Images from 47% to 97%

The research team focused on this biological principle.

In contrast, the technology developed by the research team adopts an "in-sensor image processing" approach, where the sensor detects light while simultaneously performing some image processing.

The research team solved this problem using tungsten diselenide (WSe2), an ultra-thin two-dimensional semiconductor material.

When noise-containing images were directly input into an AI trained on clean character images, recognition accuracy was only 47.7%.

However, when the same images were input after applying the newly developed image processing, accuracy jumped to 97.3%—more than double.