Brain-Inspired Vision System Reconstructs Images Through Dense Fog

Researchers at Brown University School of Engineering have developed a brain-inspired imaging system capable of detecting and tracking moving objects through dense fog and turbid water. The breakthrough, detailed in a 2026 study published in Advanced Science, introduces an end-to-end neuromorphic architecture that effectively filters out light scattering to reconstruct clear visual data in obstructed environments. The system operates by combining a dynamic vision sensor with a spiking neural network. Unlike conventional cameras that capture static full frames, the event-based sensor records only brightness changes at the pixel level, firing asynchronous data pulses when motion or edges are detected. This approach mimics the human retina, which prioritizes visual changes over continuous imagery. Because scattered light from fog or water shifts slowly compared to a moving target, the sensor naturally suppresses background noise while isolating the object. The resulting sparse spike trains are then processed by a neuromorphic computer model that emulates the visual cortex. This deep spiking neural network reconstructs the shape of the obscured target and calculates its trajectory in fractions of a second. Laboratory validation demonstrated the system’s efficacy in challenging optical conditions. Testing inside controlled fog chambers and turbid water tanks revealed that the reconstructed images achieved structural similarity scores of up to 96 percent compared to the original targets. The technology successfully tracked randomly moving alphanumeric characters and bird silhouettes, performance levels that standard optical systems and human vision cannot match under identical scattering conditions. Additionally, the architecture offers significant energy efficiency. The sensor consumes only tens of milliwatts, and the spiking network drastically reduces computational load by processing data exclusively when relevant spikes are generated, avoiding the power demands of dense frame processing. The research team, led by postdoctoral researcher Ning Zhang and senior author Arto Nurmikko, envisions immediate deployment in environments where light scattering severely compromises optical visibility. Potential applications include enhanced perception systems for autonomous vehicles, search-and-rescue drones operating during wildfires or storms, and underwater navigation platforms. The team also notes prospective utility in specialized medical imaging contexts. Current limitations remain tied to the event-based design. Because the sensor relies on brightness variation, stationary objects remain undetectable, and the network currently outputs silhouette reconstructions rather than full grayscale imagery. Sensitivity also diminishes in low-light environments. To address these constraints, the researchers are developing a light-intensifier front end to improve performance in dark conditions and exploring depth-resolved methodologies, such as time-of-flight measurements and dual-sensor stereoscopic configurations, to expand the system into three-dimensional target tracking. By merging bio-inspired sensing with neuromorphic computing, the Brown University team has established a scalable framework for penetrating visually obstructed media, with ongoing work aimed at broadening operational parameters for next-generation autonomous and diagnostic systems.