Programmable vision chip enabling, High fram rate and low latency image analysis
What is Retine?
- In collaboration with CEA-List for embedded imaging processing, CEA-Leti has developped RETINE. Thanks to 3D IC stacking technology, it combines an image sensor and a parallel array of processors to form a single vision circuit. RETINE can hence acquire upto 5500fps while featuring flexible scene analysis capabilities; It results in downto millseconds latency between acquisition and computation outputs and the resulting data stream can be reduced to the minimal relevant amount.
- RETINE's processing capabilities allow to implement pixel based local processing very well, from low-level image processing such as image transformation or filtering to medium-level image processing such as feature detection and extraction at high speed. These types of algorithms benefit directly from the fine-grained parallelism and the local communication mechanisms provided by the RETINE chip. The design of this circuit is really scalable as it is built by replicating a 3D elementary brick composed of a matrix of pixels and an elementary processor. Thus, by increasing the resolution, the associated computing power increases in an equivalent way.
Applications:
RETINE helps improve:
- Robotics: compact solution for positioning or tracking control
- Quality inspection in manufacturing leveraging on the fly optical monitoring
- High-speed optical monitoring
- Safety with rapid presence detection to trigger safety devices or promptly cut off engines
- Drones: vision chips are helpful for flight assistance through ground positioning and speed estimation, field analysis or ground modeling.
- Implement machine learning algorithm inside the sensor to enhance image with semantic information
What's new?
- RETINE vision chip is built around a scalable array parallel structure.The 3D stacked implementation brings a high data bandwidth from the sensor to the processing elements, allowing high speed, minimum latency, image analysis.
- The 192 multi-core processors enable the execution of differentiated code based on a specific area of the image, allowing multi-processing in the same focal plane array.
- Depending on the processing architecture, the chip allows different local feedback on pixel acquisition (integration time, power on/off, etc.). As a result the chip can be seen as a flexible system able to significantly reduce the need of communication to a main external driver element.
What's next?
- RETINE provides a major breakthrough in terms of flexible computing with low latency. It will address the growing needs for image analysis in various industries. In close collaboration between CEA-Leti and CEA-List institutes, the next generation of Retine is under development to reduce pixel size and integrate more processing power into the sensor.