Small sonar targets are not just hard to see, they are easy to misinterpret, and in ocean observation, misinterpretation carries real consequences. The method described here, a YOLOv26n detector enhanced with Gated Output-16 and MPDIoU-NWD loss integration, represents a measurable step forward in how we extract reliable signals from noisy, low-contrast seafloor data. The numbers speak clearly: a 2.22 percentage-point improvement in mAP50:95 and a 4.22-point gain in recall, all while keeping the model lightweight at 2.5 million parameters and running at 107 frames per second. This is not a speculative advance; it is an empirically validated refinement of a detection pipeline that prioritizes the smallest, most ambiguous objects, precisely the ones that get missed in routine surveys.
What matters most here is the architectural decision behind the improvement. Gated Output-16, placed at the P3 detection head, starts with gates at zero, preserving the original feature path at initialization. That design choice avoids the common pitfall of over-correcting a model before it has learned what it needs to see. The scale-aware supervision, MPDIoU for ordinary targets, Normalized Wasserstein Distance for those under 1024 pixels², treats small objects as a distinct class of problem, not an afterthought. This mirrors the logic we see in other validated applications across our publication, such as the Validated data and selective gear drive sustainable crab harvesting approach, where precision at the level of individual animals depends on the same principle: you cannot manage what you cannot measure. Similarly, the Validated vision model counts deformable shrimp larvae with measurable accuracy work demonstrates that robust detection under variable conditions requires loss functions and architectures designed for the specific morphology of the target. Here, the target morphology is the sonar signature itself, blurred, small, embedded in complex backgrounds, and the gating mechanism ensures the model does not erase useful information before it has been evaluated.
For practitioners working with side-scan sonar data, the practical takeaway is direct: you can deploy a model with 2.5 million parameters that processes an image in 9.3 milliseconds and catches 85.77% of targets, compared to 81.55% with the baseline. That 4.22-point recall gain means four more detections out of every hundred targets that were previously missed. In applications ranging from mine detection to habitat mapping to infrastructure inspection, those missed targets are not statistical noise, they are the difference between a complete survey and a partial one. The model runs under a unified FP16 protocol, meaning it is deployable on edge hardware without sacrificing precision. And because the improvement is driven by a targeted architectural modification rather than brute-force scaling, it suggests that similar gains are possible in other detection tasks where small, weak features dominate.
One open question worth watching: how does the gated output generalize to other sonar frequencies, water column conditions, or target classes beyond the three in this dataset? The baseline and final model were evaluated on a single dataset with 1,095 training images. The recall gain is substantial, but the validation set is not enormous. The next step should be a cross-dataset evaluation, testing the same architecture on a different sonar collection, ideally from a different geographic region or sensor platform. If the recall improvement holds under those conditions, then Gated Output-16 becomes a candidate for integration into operational survey pipelines. Until then, it is a well-validated laboratory result with a clear path toward deployment, and that is exactly the kind of incremental, reliable progress that moves ocean intelligence forward.