Video Capturing and Fish Recognition Module

With underwater HD cameras deployed in the rivers, the video streams are fed to an edge device. The edge devices are physically close to the rivers, with the benefit of reduced latency. A highly optimized AI inference software is implemented for real-time fish recognition.

By presenting the AI module with tens of thousands of labeled pictures, the algorithm quickly comes up to desired level of accuracy. As more and more species of fish passes through the system, the algorithm learns even more and becomes better and better

Having the edge device close to the sorting mechanism proves to be highly efficient. Currently the species of fish is recognized by the algorithm within 8 milliseconds. Different light and background colors impact the capabilities of fish recognition. Light colored background, but not completely white proves to be best. Midnight sun has a positive effect, making the recognition able to run 24/7