by The continual progress in machine learning methods and sensing technologies in recent years has brought about new opportunities in the field of object detection and tracking. The precise and automated identification of visual targets, which is also referred to as intelligent machine vision, holds significant potential in a wide array of applications, from bolstering security systems and surveillance tools to monitoring the environment and analyzing medical imaging data.
However, traditional sensors based on complementary metal-oxide-semiconductor (CMOS) technology often struggle to accurately detect and track dim targets in challenging lighting conditions or low visibility settings, as they may have limitations in extracting essential features like edges and corners from images.
To address this limitation, a team of researchers from Nanjing University and the Chinese Academy of Sciences has introduced a pioneering approach to enhance sensors’ ability to detect dim targets in complex environments. Their groundbreaking method, as detailed in a recent publication in Nature Electronics, involves the implementation of in-sensor dynamic computing, which integrates sensing and processing functionalities within a single device.
In low-contrast optical environments, accurately discerning weak targets has been a longstanding challenge due to issues such as low accuracy and poor robustness. Shi-Jun Liang, the senior author of the research, highlighted that the small intensity disparity between target and background light signals, with the target signal often being overshadowed by background noise, exacerbates the difficulty of intelligent target perception.
Traditional techniques for static pixel-independent photoelectric target detection typically rely on CMOS-based sensors, which may struggle to distinguish target signals from background signals effectively. To overcome these challenges, researchers have been exploring innovative hardware development principles utilizing low-dimensional materials to enhance sensors’ robustness and precision in low-contrast optical environments.
The team’s approach leverages multi-terminal photoelectric devices based on graphene/germanium mixed-dimensional heterostructures to enable dynamic control of correlation strength between neighboring devices in the optoelectronic sensor. By dynamically modulating convolution kernel weights based on local image intensity gradients, the in-sensor dynamic computing units can adapt to varying image content, providing a more sophisticated and adaptable solution for target detection and tracking in complex lighting situations.
The researchers’ novel paradigm introduces dynamic feedback control between interconnected and neighboring optoelectronic devices, paving the way for ultra-accurate and robust recognition of contrast-varying targets under unfavorable lighting conditions. Unlike conventional sensors where devices operate independently, the correlated operation in the in-sensor dynamic computing technology enhances the sensor’s ability to detect and track dim targets effectively.
The proposed approach’s compatibility with conventional CMOS technology and scalability potential for large-scale on-chip integration showcase its promise for real-world applications. In the future, the team plans to explore the expansion of detection wavelengths to near-infrared or mid-infrared bands to broaden the applicability of the technology in various low-contrast scenarios such as remote sensing, medical imaging, monitoring, security, and early-warning systems in low visibility meteorological conditions.
*Note:
1. Source: Coherent Market Insights, Public sources, Desk research
2. We have leveraged AI tools to mine information and compile it
