Infrared cameras are often used for low-light target monitoring, but in aquaculture, water absorption of infrared radiation prevents their use for imaging beneath the water surface. To resolve this issue, a fast modeling mechanism dedicated to fish segmentation tasks in low-light underwater settings was presented, which integrates a 2MP(Mega Pixels) underwater low-light adaptive camera with a knowledge distillation framework based on a large segmentation model. For the low-light camera, an imaging algorithm optimized for NPU (Neural Processing Unit) chips significantly enhances the performance in underwater monitoring scenarios. For the fish segmentation model, the proposed lightweight knowledge distillation framework integrating the Segment Anything model and YOLOv8 eliminates the need for manual mask annotation while maintaining accurate segmentation performance. Comparative experiments show that the optimized imaging algorithm enables low-light cameras to achieve superior imaging quality under low-light and dark conditions compared to other methods, and when combined with the rapidly constructed fish segmentation model, it can accurately and reliably detect and segment targets.
Zhao et al. (Sun,) studied this question.