Abstract
The shrimp trawl fishery in Norway struggles with catch efficiency, sustainability and profitability. This is largely due to great search efforts, high fuel prices and a rising carbon tax. A widespread weakness in today's shrimp trawling is the lack of real-time information on catches while trawling. A plethora of systems are being developed that aim to provide real-time catch information as decision support. One of these is the CamSounder trawl camera system, developed by Scantrol Deep Vision. The CamSounder is an underwater stereo camera system for deployment in the trawl, with onboard/edge processing power for running AI models, and an acoustic link for wireless transmission to the bridge.
We will present our methodology and results from developing real-time shrimp and bycatch detection models for the CamSounder. Our benchmark includes models based on the RF-Detr, DEIMv2, and YOLO architectures. These were developed based on images collected from eight surveys carried out with MS Vollerosa in May and September 2025 and February 2026. These images were semi-automatically annotated using an annotation model bootstrapping procedure which we will explain. Moreover, half of the surveys were carried out with a greyscale sensor on the left channel instead of a colour sensor, to experiment with a sensor having more light sensitivity. This allows us to benchmark the sensors against each other, and our preliminary results indicate improved performance for the greyscale sensor.
Finally, we present a general algorithm for matching detections across stereo channels. This lets us avoid double counting and provide shrimp candidates for size estimation, from which catch statistics are computed. We will demonstrate the functionality using previously collected data, and discuss the potential for further improvements.
This activity was funded by the FHF through the project "Sanntidsovervåkningteknologi for beslutningsstøtte i trålfiske" (project no. 901976).