Analysis of Methods to Improve Underwater Object Detection in Various Environments
Nico Katz
Affiliation: La Jolla Country Day School, La Jolla, CA, USA
IJSCAR Vol. 3, Issue 4 (2026) · pp. 44–56
Abstract
We evaluate how single-augmentation and aggregate-augmentation training affect the robustness of a Faster R-CNN detector on an underwater target-detection task, and we assess a physics-based dehazing stage built on ten synthetic attenuation levels ordered as the Jerlov water types are. Seven detectors, one of them a control, were tested on a clean split, a composite degradation, each degradation alone, and four degradations held out of every training policy. All seven fall within 0.805–0.841 AP on clean data, so clean validation does not discriminate between augmentation policies. The aggregate model’s lead over the control is the same on trained-on and held-out degradations (0.026 AP each) and reappeared in two retrained seeds per policy (0.019 and 0.021 AP), though two seeds cannot establish reproducibility; its composite-condition gap (0.186 AP) did not reproduce on retraining (0.406 and 0.412). The aggregate is not the best of the seven on held-out degradations, an ordering that includes checkpoints predating per-worker data-loader seeding, and its advantage is concentrated in two blur-like conditions on the validation split (83%) but not on a held-back 97-image partition (44%). The physics inversion alone raises AP from 0.691 to 0.819 across three conditions and beats a Dark Channel Prior baseline, whereas adding contrast enhancement and white balance lowers it to 0.680; this comparison favors our method by construction, since the inversion is given the true attenuation level and path length, the frames are generated by the equation it inverts, and the coefficients are our own. Across the full sweep, which used the three-stage pipeline, correction gains 0.197 AP in coastal water and costs 0.117 AP in open ocean. Augmentation and correction approximately add for the aggregate policy but overlap for the other five. Clean validation performance may not predict robustness under degraded conditions.
Keywords: