Smart Image Sensor

Wireless image sensors are integral components of many critical applications such as surveillance, remote monitoring, intelligence gathering, unmanned vehicles, to name a few. We are interested in exploring fundamental research concepts on intelligent computing, energy-efficiency, and security, to create smart and autonomous image sensors. Our goal is to explore smart sensors at different operating specifications, ranging from high speed high resolution sensors to self-powered low frame rate sensors. The key contributions in this field are: 

  1. Smart Sensors with Integrated Machine Learning.
    • Adaptive Image Sensor with Embedded Machine Learning: We are exploring integration of deep learning in an image sensor to enhance the information quality delivered by a sensor. The integrated deep learning is used to enable context-defined data reduction techniques in the sensor platform. Moreover, the embedded machine learning is used in a feedback path to guide the sensors to collect data with maximum information content. In short, we are building cameras that sense information not signal.
    • Hardware Architecture for Sensors with Integrated Learning. We are exploring system architecture for high-performance smart sensors with integrated deep learning accelerators. 3D stacking of sensors, read-out circuits, memory, and computing logic is being explored as a system integration approach to achieve this goal. 
  2. Lightweight Energy-Neutral Image Sensors for Remote Surveillance 
    • Self-powered Image Sensor: We are exploring design of image sensors where the on-chip pixel-array can be dual-purposed to both sense image as well as harvest energy. The design of innovative pixel arrays coupled with efficient power conversion/regulation and lightweight image processing are being explored to create a truly self-powered image sensor for remote surveillance applications.
    • Lightweight Algorithms for Complex Vision Processing: We are  innovating lightweight algorithms for vision processing tasks for resource-/energy- constrained platforms. A key goal is to explore algorithmic approaches to reduce memory demand of conventional as well as deep learning based image/vision processing algorithms. 
  3. Environment Adaptive Sensors with Real-time Control: The sensors operate in uncontrolled environment with potentially high variations in input quality, communication bandwidth, or available energy. We are developing real-time control algorithms to maintain quality-of-serive under varying environment by simultaneously tuning vision algorithms and sensor hardware. The collaboration between sensors and host is used to enable higher information quality under energy and bandwidth constraints.