Page 41: of Marine Technology Magazine (November 2025)

sumer hardware with limited energy, as the system runs on battery power: a compromise on performance was therefore made with this hardware.

The detection model we used in 2020 is YOLOv4 (You Only

Look Once), a deep learning algorithm designed for real-time object detection. The model was adapted to take full advan- tage of the VPU's capabilities. We used transfer learning to adjust the model parameters to our dataset.

Creating a dataset and training the model

To train our specialized model for Pterois Volitans/Miles,

A lionf sh in its or lion? sh, a speci? c dataset was compiled. This dataset natural habitat. © Adobe Stock/crisod was assembled from various sources, including open-access resources and recordings made in natural environments and aquariums. This diversity of sources provided a broad repre- sentation of different lighting conditions and viewing angles, reinforcing the robustness of the future model.

The creation of this dataset was a crucial step in ensuring the relevance and performance of the model. By choosing to target a single species exclusively, we signi? cantly reduced the risk of false positives (detection errors with other species), while optimizing inference time and, indirectly, the energy consumption of the device.

Detector test

The prototype was tested in March 2021 under near-real con- ditions, with budget in mind: submerged in a tank containing lion? sh belonging to the Océanopolis marine center, with a return via the mobile telephone network enabling detection alerts and snapshots remote transmission.

Maximum performance with this hardware architecture en- abled real-time processing at approximately 5 FPS (frames per second). The ultimate goal is to achieve 1 FPS in order to ? nd a better compromise in terms of power consumption, as this species moves slowly.

The detector makes a local copy of the full-resolution photos of the detections made, which will ultimately enrich our learn- ing dataset. It could also send the number of detections via an

IoT network (e.g. LoRa), where bandwidth is very limited.

Power consumption averages less than ? ve watts at 5 FPS,

SYRENE V1 ready to be vertically and its battery life exceeds 100 hours without optimization for immersed with its mirror. the available battery volume in the enclosure.

Credit: IFREMER S. Barbot Credit: IFREMER S. Barbot

Lionf shes’ tank with

SYRENE in the right corner. www.marinetechnologynews.com 41

MTR #8 (34-47).indd 41 MTR #8 (34-47).indd 41 11/20/2025 5:50:11 PM11/20/2025 5:50:11 PM