What you need to know
- Measure data and energy for normal operation and exceptions.
- Preserve enough image evidence to investigate uncertain results.
Processing closer to the camera
Bharat Semi raised $24 million in seed financing, with Info Edge Group, South Park Commons and Indus Bridge Ventures leading the round. The company describes its work as building image sensors for applications such as machine vision and commercial drones, with processing at the edge. The financing supports development of programmable sensors that combine imaging and AI.
Decide what the rest of the device needs
Imagine a camera on a packaging line that usually needs to transmit only the location of a misplaced label. On an uncertain reading, a technician may need the original image to understand whether the label was folded or the lens was dirty. A design that sends only the inferred location could save bandwidth while making those exceptions harder to diagnose.
The device maker can design for both cases by specifying a normal output and an exception output. Evaluation would then count the bytes transferred and the energy used per inspected package, including the images retained for review. A comparison with a separate processor should use the same lighting, line speed and error criteria.
Allow for a change in the inspection task
A line that starts with label placement may later need to read a lot number. That could change the resolution, processing time and raw data needed. A prototype review should include at least one plausible change in the task so the buyer can see which parts of the system need new software and which need different hardware.
The technical opportunity is to send less data without losing the evidence an operator needs. Buyers can judge this at the level of a complete device as prototype designs become available.



