Machine learning evaluation demonstrates reduced counting error in crowded aquaculture imagery, indicating that high-density loss weighting enhances monitoring without inference overhead.
Key Points
Evaluate whether high-density-aware density modeling can reduce underwater fish counting errors in low-visibility, high-density recirculating aquaculture without adding inference complexity.
Evaluated an annotated 900-image dataset (TV5/IOC) using three seed-controlled splits of 720 training and 180 validation single-view images.
Implemented high-density-aware density modeling (HDA-DM), which assigns a mild sample weight solely to the density loss of high-count training frames.
Ablated against cross-view density distillation (CVDD) and compared density-only weighting to joint density, count, and shape loss supervision.
HDA-DM reduced three-seed mean MAE on the Main900 dataset from 4.024 ± 0.578 to 3.739 ± 0.414, representing a 7.07% relative error reduction.
HDA-DM lowered the mean MAE on the challenging Hard200 subset from 9.799 ± 1.792 to 8.863 ± 0.304, achieving a 9.56% error reduction.
Adding cross-view density distillation did not enhance performance (Main900 MAE: 3.987 ± 0.635; Hard200 MAE: 9.666 ± 1.373), and HDA-DM required zero extra parameters or operations at inference.