LoRA Enhanced Contrastive Learning with SAS Vision Transformers

# LoRA Enhanced Contrastive Learning with SAS Vision Transformers
Source: [https://arxiv.org/html/2609.21061](https://arxiv.org/html/2609.21061)
## LoRA Enhanced Contrastive Learning with SAS Vision TransformersThanks:This work was supported by the Office of Naval Research \(N0001426GI00772\) and their Internal Applied Research program\.

Dan ZimmermanAffiliation:Center for Connected Autonomy & AI Florida Atlantic University Boca Raton, USA dzimmerman2021@fau\.eduFrank E\. Bobe IIIAffiliation:Naval Surface Warfare Center Panama City Division Panama City, USA frank\.e\.bobe\.civ@us\.navy\.milAmelia L\. McCormackAffiliation:Department of Computer Science Florida State University Tallahassee, USA alm20h@fsu\.eduAffiliation:Matthew CookAffiliation:Naval Surface Warfare Center Panama City Division Panama City, USA matthew\.g\.cook12\.civ@us\.navy\.milGregory D\. VetawAffiliation:Naval Surface Warfare Center Panama City Division Panama City, USA gregory\.d\.vetaw\.civ@us\.navy\.milAffiliation:

###### Abstract

Automatic target recognition \(ATR\) with synthetic aperture sonar \(SAS\) enables advanced naval capabilities, but deep learning approaches remain constrained by human\-in\-the\-loop assessment and by the extreme scarcity of imaged targets against background clutter\. We present a three\-stage parameter\-efficient adaptation framework with contrastive refinement, adapting self\-DIstillation with NO labels v3 \(DINOv3\) Vision Transformer \(ViT\) models to underwater SAS ATR\. Stage 1 applies a Low\-Rank Adaptation \(LoRA\) workflow to bridge the gap between pretraining on natural images and the physics of underwater acoustic propagation while freezing the ViT backbone\. Stage 2 implements a “Refiner” using hard negative mining to harden the decision boundary against acoustic mimics: naturally occurring seafloor features, such as rock outcrops and sediment formations, whose sonar signatures resemble man\-made targets\. Stage 3 applies Supervised Contrastive Learning \(SupCon\) to pull targets into a compact manifold away from clutter\. We evaluate on at\-sea SAS data under a mission\-level geographic split, read every arm at a matched85%85\\%test recall, and repeat each comparison over three random seeds\. Adaptation accounts for the entire effect: LoRA raises the Area Under the Precision\-Recall Curve \(AUPRC\) from0\.3000\.300to0\.679±0\.0270\.679\\pm 0\.027over the identical frozen backbone, a\+0\.38\+0\.38gap an order of magnitude beyond any other effect we measure, and rank44reaches the same result while training only0\.26%0\.26\\%of its weights\. Neither refinement stage separates from its matched control: mining is worth−0\.0045±0\.0119\-0\.0045\\pm 0\.0119AUPRC against a random curriculum of equal size, and SupCon\+0\.0002±0\.0096\+0\.0002\\pm 0\.0096against the stage it refines\. We report both null results with their mechanisms, namely that the curriculum is mined on a split the encoder has already fit, and that the supervised stages impose most of the target–clutter geometry before the contrastive term is applied\. One efficient adaptation stage is sufficient; the curriculum stacked on top of it is not\.

![Refer to caption](https://arxiv.org/html/2609.21061v1/Figures/lora1.png)Fig\. 1:The parameter\-efficient fine\-tuning \(PEFT\) architecture used for adapting the DINOv3 ViT\-L model\. Dual\-band \(High\-Frequency \(HF\) and Broadband \(BB\)\) SAS imagery is passed through a trainable 2\-to\-3 channel stem\. The frozen ViT\-L backbone is augmented with trainable LoRA modules in its linear projection layers\.## IIntroduction

Automatic target recognition \(ATR\) enables advanced naval capabilities for underwater and maritime applications\[[21](https://arxiv.org/html/2609.21061#bib.bib16),[24](https://arxiv.org/html/2609.21061#bib.bib14),[4](https://arxiv.org/html/2609.21061#bib.bib19),[28](https://arxiv.org/html/2609.21061#bib.bib11)\], including identification of objects that endanger littoral operations\[[8](https://arxiv.org/html/2609.21061#bib.bib15),[25](https://arxiv.org/html/2609.21061#bib.bib17),[26](https://arxiv.org/html/2609.21061#bib.bib18)\]\. Supervised deep learning has advanced synthetic aperture sonar \(SAS\) analysis, but it relies on massive human\-verified datasets, and in operational environments target imagery is scarce against a vast background of clutter\. Much of that clutter consists of*acoustic mimics*: naturally occurring seafloor features, such as rock outcrops, boulders and sediment formations, whose highlight\-and\-shadow signatures resemble those of man\-made targets\. These features degrade standard classifiers and compound the underlying class imbalance\[[29](https://arxiv.org/html/2609.21061#bib.bib12)\]\.

To bypass this bottleneck we introduce a parameter\-efficient domain adaptation framework for SAS ATR\. The framework leverages a DINOv3\[[20](https://arxiv.org/html/2609.21061#bib.bib1)\]Vision Transformer \(ViT\-L\) foundation model pretrained on natural imagery, which does not capture the acoustic propagation physics of sonar\. To bridge the gap between natural\-image pretraining and the physics underlying underwater acoustic propagation, we propose a three\-stage Low\-Rank Adaptation \(LoRA\)\[[6](https://arxiv.org/html/2609.21061#bib.bib6)\]workflow comprising: \(1\) an acoustic domain adaptation stage that aligns the ViT\-L with foundational sonar features; \(2\) an active\-learning “Refiner” that uses hard negative mining to map the decision boundaries of complex seafloor textures; and \(3\) a Supervised Contrastive Learning \(SupCon\)\[[7](https://arxiv.org/html/2609.21061#bib.bib7)\]stage that projects the data into a separable geometric feature space\.

Crucially, we evaluate each stage against a*matched control*: a run identical in curriculum size, optimizer, schedule and weight initialization, differing only in the single mechanism that stage contributes\. Comparing a stage against its own control, rather than against the final end\-to\-end accuracy of the assembled pipeline, isolates the effect of the mechanism from the effect of longer training\. Every arm is read at a matched 85% test recall, and we repeat every close comparison across three random seeds\.

This protocol changes the conclusion\. Adapting the frozen backbone with LoRA raises the Area Under the Precision\-Recall Curve \(AUPRC\) by more than a factor of two while training at most4\.2%4\.2\\%of the backbone’s weights, and as little as0\.26%0\.26\\%suffices; neither subsequent stage separates from its control\. We report those two null results in full rather than omitting them, because each is accompanied by a mechanism that explains it: the mined curriculum is selected on a split the encoder has already fit, and the supervised stages impose most of the target–clutter geometry before the contrastive term is applied\. The practical consequence is that engineering effort belongs in the single adaptation step rather than in the multi\-stage curriculum built on top of it\.

Our contributions are: \(i\) a parameter\-efficient adaptation of DINOv3 to dual\-band SAS recovering a\+0\.38\+0\.38AUPRC gap over the same frozen backbone; \(ii\) evidence that adapter capacity saturates at very low rank and that low\-rank adaptation is not separable from full fine\-tuning at either backbone scale we tested; and \(iii\) a controlled, matched\-control accounting of two widely assumed pipeline components, hard negative mining and supervised contrastive refinement, establishing that neither repays its cost at this data scale and supplying the mechanism behind each null result\.

## IIRelated Work

Deep learning for SAS ATR\.SAS ATR has relied on supervised convolutional neural networks \(CNNs\)\[[5](https://arxiv.org/html/2609.21061#bib.bib4)\]to extract spatial features from acoustic imagery\[[26](https://arxiv.org/html/2609.21061#bib.bib18),[21](https://arxiv.org/html/2609.21061#bib.bib16)\], extended to spectral partitioning\[[25](https://arxiv.org/html/2609.21061#bib.bib17)\], volumetric normalization\[[24](https://arxiv.org/html/2609.21061#bib.bib14)\], and seafloor classification\[[8](https://arxiv.org/html/2609.21061#bib.bib15)\]\. Lightweight detectors, including optimized You Only Look Once \(YOLO\) variants\[[29](https://arxiv.org/html/2609.21061#bib.bib12),[27](https://arxiv.org/html/2609.21061#bib.bib13)\]and few\-shot transformer detectors\[[28](https://arxiv.org/html/2609.21061#bib.bib11)\], have advanced real\-time analysis\. The extreme imbalance between targets and clutter nonetheless remains a persistent challenge\.

Foundation models and parameter\-efficient adaptation\.Vision Transformers\[[3](https://arxiv.org/html/2609.21061#bib.bib3)\]pretrained with the DINOv2\[[15](https://arxiv.org/html/2609.21061#bib.bib2)\]and DINOv3\[[20](https://arxiv.org/html/2609.21061#bib.bib1)\]self\-supervised frameworks generalize strongly across optical imagery\[[19](https://arxiv.org/html/2609.21061#bib.bib5)\], but the shift to underwater acoustic propagation is challenging\. Full fine\-tuning at this size is prohibitive without distributed optimizations\[[17](https://arxiv.org/html/2609.21061#bib.bib9),[16](https://arxiv.org/html/2609.21061#bib.bib10)\]\. Parameter\-efficient fine\-tuning\[[13](https://arxiv.org/html/2609.21061#bib.bib8)\]has emerged as an effective strategy for adapting large ViTs to specialized domains such as SAS ATR\. LoRA\[[6](https://arxiv.org/html/2609.21061#bib.bib6)\]freezes the pretrained weights and injects trainable low\-rank updates that substantially reduce the computational budget\. Later variants improve its flexibility by training nested ranks\[[11](https://arxiv.org/html/2609.21061#bib.bib28)\]or adjusting the rank during training\[[22](https://arxiv.org/html/2609.21061#bib.bib27)\]rather than fixing it in advance\.

Contrastive learning and extreme class imbalance\.Re\-sampling\[[1](https://arxiv.org/html/2609.21061#bib.bib23)\]and cost\-sensitive objectives\[[12](https://arxiv.org/html/2609.21061#bib.bib22)\]are limited at sonar imbalance ratios\. Contrastive frameworks\[[2](https://arxiv.org/html/2609.21061#bib.bib21),[10](https://arxiv.org/html/2609.21061#bib.bib20)\]instead maximize agreement between augmented views, and SupCon\[[7](https://arxiv.org/html/2609.21061#bib.bib7)\]extends this with labels\. Pairing SupCon with hard\-negative mining\[[9](https://arxiv.org/html/2609.21061#bib.bib24)\]can enforce separability that cross\-entropy struggles to achieve\. This work builds on an earlier investigation of ViT\-L backbones and contrastive loss for SAS\[[23](https://arxiv.org/html/2609.21061#bib.bib26)\], and reports the Matthews correlation coefficient \(MCC\)\[[14](https://arxiv.org/html/2609.21061#bib.bib25)\]as its imbalanced\-data metric\.

## IIIMethodology

We adapt the DINOv3 ViT for SAS ATR in athree\-stage pipeline, moving from broad feature alignment to adversarial refinement and finally to geometric contrastive regularization\.

### III\-AStage 1: Foundational LoRA Domain Adaptation

To bridge the domain gap between pretraining a DINOv3 ViT\-L backbone on massive amounts of optical imagery and dual\-band acoustic physics, we implement the parameter\-efficient fine\-tuning \(PEFT\) architecture of Fig\.[1](https://arxiv.org/html/2609.21061#S0.F1)\. Because the backbone expects three channels, a trainable1×11\\times 1convolutional 2\-to\-3 channel “stem” projects the dual\-band high frequency \(HF\) and broadband \(BB\) SAS magnitude channels into a compatible space\. Trainable Low\-Rank Adaptation \(LoRA\) modules are then injected into the query \(QQ\), key \(KK\), value \(VV\), and output \(OO\) linear projections of each self\-attention block\. The placement is*attention\-only*; that is, the feed\-forward projections carry no adapter, so every result here is obtained by re\-weighting attention alone\.

This architecture is trained in two ways that differ only in the objective and in the role of the Multilayer Perceptron \(MLP\)\.The supervised form is the pipeline’s Stage 1 throughout this paper: it is what Stages 2 and 3 build upon and what appears in the Stage 1 rows of Table[I](https://arxiv.org/html/2609.21061#S5.T1)\.There, the MLP is a binary classifier over the final Classify token \(CLS\) embedding, trained with the Asymmetric Loss of Ridnik et al\.\[[18](https://arxiv.org/html/2609.21061#bib.bib29)\]over class\-balanced batches, so labels enter both the objective and the sampler, and Stage 2 initializes from this encoder\.

The self\-supervised form is a separate comparison module, which is never used as an input to a later stage\. The same MLP acts instead as a projection head onto a 128\-dimensional unit hypersphere optimized with the Contrastive Clustering objective\[[10](https://arxiv.org/html/2609.21061#bib.bib20)\]over two augmented views\. That objective is not purely instance\-level\. It combines an instance Information Noise\-Contrastive Estimation \(InfoNCE\) term, as in Simple Framework for Contrastive Learning of Visual Representations \(SimCLR\), with a cluster\-assignment head \(K=50K=50\) grouping the batch without labels, which are used only for sampling\. Its frozen features are scored by a linear probe, appearing as the Contrastive Clustering Self\-Supervised Learning \(CC\-SSL\) row of Table[I](https://arxiv.org/html/2609.21061#S5.T1)\.

### III\-BStage 2: Adversarial Refinement

Performance after Stage 1 is limited by geometrically complex clutter\. To harden the decision boundary against these “acoustic mimics,” we freeze the Stage 1 encoder, extract embeddings for the training set, train a lightweight linear probe on those frozen features to score each sample, and take the topK=5,000K=5\{,\}000clutter samples by false\-positive confidence as hard negatives\. We then build a curriculum of all known target signatures plus those negatives, initializing from the Stage 1 adapters and band stem rather than at random so the network continues adapting its attention subspace, and fine\-tune at2×10−52\\times 10^\{\-5\}with the same Asymmetric Loss\[[18](https://arxiv.org/html/2609.21061#bib.bib29)\]\. Because this curriculum alters the class balance regardless of*which*clutter is chosen, mining confounds the informativeness of the selected negatives with simple exposure rebalancing\. We therefore pair every Stage 2 run with arandom exposure control, identical but for drawing its5,0005\{,\}000snippets uniformly rather than by score, so the difference between the arms isolates the mining signal\.

### III\-CStage 3: Contrastive Geometric Regularization

To resolve the remaining ambiguity between targets and acoustic mimics, we introduce a Supervised Contrastive Learning \(SupCon\) stage\[[7](https://arxiv.org/html/2609.21061#bib.bib7)\], treating the embedding space as a geometric problem that structurally enforces class separability\. Initializing from the refined Stage 2 adapters, stem and classifier head, we continue training on the hard\-negative curriculum with the dual\-loss objectiveℒtotal=λsup​ℒsup\+λcls​ℒcls\\mathcal\{L\}\_\{\\text\{total\}\}=\\lambda\_\{\\text\{sup\}\}\\mathcal\{L\}\_\{\\text\{sup\}\}\+\\lambda\_\{\\text\{cls\}\}\\mathcal\{L\}\_\{\\text\{cls\}\}, whereℒcls\\mathcal\{L\}\_\{\\text\{cls\}\}is the imbalance\-aware asymmetric loss over the classification logits of two stochastically augmented views per sample, andℒsup\\mathcal\{L\}\_\{\\text\{sup\}\}is the Supervised Contrastive loss\[[7](https://arxiv.org/html/2609.21061#bib.bib7)\]\. A temporary projection head maps representations to a 128\-dimensional unit hypersphere, and for a batch ofNNsamples stochastically doubled to a multiview batch of2​N2Nviews,

ℒsup=∑i∈I−1\|P⁡\(i\)\|​∑p∈P⁡\(i\)log⁡exp⁡\(zi⋅zp/τ\)∑a∈A⁡\(i\)exp⁡\(zi⋅za/τ\),\\mathcal\{L\}\_\{\\text\{sup\}\}=\\sum\_\{i\\in I\}\\frac\{\-1\}\{\|P\(i\)\|\}\\sum\_\{p\\in P\(i\)\}\\log\\frac\{\\exp\(z\_\{i\}\\cdot z\_\{p\}/\\tau\)\}\{\\sum\_\{a\\in A\(i\)\}\\exp\(z\_\{i\}\\cdot z\_\{a\}/\\tau\)\},\(1\)withI≡\{1,…,2​N\}I\\equiv\\\{1,\\dots,2N\\\}indexing the multiview batch,ziz\_\{i\}the projected anchor,P⁡\(i\)P\(i\)the positive views sharing its class label,A⁡\(i\)≡I∖\{i\}A\(i\)\\equiv I\\setminus\\\{i\\\}is the index set of all other views in the batch, andτ\\tauis the temperature parameter \(set to0\.070\.07\)\. Crucially for the analysis that follows,ℒsup\\mathcal\{L\}\_\{\\text\{sup\}\}is not minimized at zero: for a perfectly class\-collapsed embedding it attains𝔼i​\[log⁡\|P​\(i\)\|\]\\mathbb\{E\}\_\{i\}\[\\log\|P\(i\)\|\], a constant fixed by batch composition, so its distance above that floor, not its raw value, indicates how much geometric structure the objective still has to impose\.

## IVImplementation and Training Details

### IV\-ADataset and Splits

We evaluate on dual\-band \(High\-Frequency and Broadband\) SAS magnitude imagery from at\-sea deployments, comprising 148 mission files partitioned at the*mission*level by geographic location, so no mission contributes to more than one split and the test set represents a genuine geographic shift rather than a random sample of the same surveys\. Training comprises 106 missions and 336,298 snippets containing 2,791 targets, which equates to a severe119:1119\{:\}1imbalance; validation and test contain 16 and 26 missions, at 36,327 snippets \(99:199\{:\}1\) and 67,083 snippets \(155:1155\{:\}1\)\.

### IV\-BPreprocessing and Augmentation

Imagery is bilinearly resized to224×224224\\times 224to match the DINOv3 ViT\-L input resolution and phase is discarded\. Each band is standardized independently to zero mean and unit variance, clamped at±8\\pm 8standard deviations to tame the heavy tail from bright target returns, identically at every stage\. For the contrastive phases we apply Rayleigh\-distributed speckle noise, random resized cropping, flipping, photometric jitter and Gaussian blurring \(σ∈\[0\.1,2\.0\]\\sigma\\in\[0\.1,2\.0\]\), so the loss is driven by target geometry rather than pixel similarity\.

### IV\-COptimization and Model Selection

The pipeline is implemented in PyTorch and trained using distributed data\-parallel processing across eight NVIDIA A6000 GPUs, giving a pooled batch of 1,024\. DINOv3 ViT\-L weights remain frozen throughout\. LoRA modules use rankr=64r=64withα=2​r\\alpha=2r, selected by a sweep overr∈\{2,4,8,16,32,64\}r\\in\\\{2,4,8,16,32,64\\\}on validation AUPRC\.

Stage 1uses Adam at1×10−41\\times 10^\{\-4\}, batch sizes of 128 per GPU, and the Asymmetric Loss\[[18](https://arxiv.org/html/2609.21061#bib.bib29)\]over class\-balanced batches; the self\-supervised comparison variant substitutes a Contrastive Clustering loss with instance temperatureτi=0\.5\\tau\_\{i\}=0\.5and aK=50K=50cluster head\.Stage 2fine\-tunes the 5,000\-hard\-negative curriculum with AdamW at2×10−52\\times 10^\{\-5\}over 30 epochs with cosine annealing, again under the Asymmetric Loss; the random\-selection control is identical except that its 5,000 clutter snippets are drawn uniformly\.Stage 3initializes from Stage 2 and trains dual\-view batches with AdamW at2×10−52\\times 10^\{\-5\}, SupCon temperatureτ=0\.07\\tau=0\.07and a 128\-dimensional projection head\. Both loss weights are unity,λsup=λcls=1\\lambda\_\{\\text\{sup\}\}=\\lambda\_\{\\text\{cls\}\}=1\. Sweepingλsup\\lambda\_\{\\text\{sup\}\}over\{0,0\.1,0\.3,3,10\}\\\{0,0\.1,0\.3,3,10\\\}moves blind\-test AUPRC by at most0\.00740\.0074, inside the±0\.0096\\pm 0\.0096paired seed interval, and disabling the term outright costs0\.00410\.0041\.

Model selection at every stage uses validation AUPRC, with early stopping after ten checks without improvement and the best checkpoint restored before a frozen evaluation pass; Section[V\-G](https://arxiv.org/html/2609.21061#S5.SS7)discusses the resulting selection bias\. Every comparison smaller than the seed\-to\-seed spread of roughly±0\.03\\pm 0\.03AUPRC was repeated at seeds 42, 100 and 999 and is reported as a mean with a95%95\\%confidence interval, or as a paired per\-seed difference where available\.

## VResults and Performance Analysis

All results are on an unseen test split \(67,083 snippets, 431 targets\)\. No label from the validation or test missions reaches any arm at any stage; the one qualification, concerning unlabeled imagery, is in Section[V\-G](https://arxiv.org/html/2609.21061#S5.SS7)\. Every arm shares one backbone, one split and one evaluation harness, and all results are means over three seeds \(42, 100 and 999\) with95%95\\%confidence intervals\.

### V\-AEvaluation Protocol

All threshold\-based metrics are read at arecall\-matched operating point: the threshold is placed so each model reaches 85% recall on the blind test set\. MCC andnFAn\_\{\\text\{FA\}\}are the threshold\-based metrics reported here, and we quote the False Positive Rate at that anchor \(FPR85\\text\{FPR\}\_\{85\}\) as the headline false\-positive rate, alongside the threshold\-free AUPRC\. Where a claim concerns the high\-recall regime specifically, we reportFPR95\\text\{FPR\}\_\{95\}and say so\.

TABLE I:Blind\-test performance\. Mean±\\pm95% CI over three seeds \(42, 100, 999\) for every arm\. MCC∗andnFAn\_\{\\text\{FA\}\}are read at a matched 85% test recall; AUPRC and AUROC are threshold\-free\. No entry is emboldened: the adapted arms have overlapping intervals and are not separable\.![Refer to caption](https://arxiv.org/html/2609.21061v1/Figures/pr_curves.png)Fig\. 2:Precision\-recall on the blind test split, seed 42\. The dotted rule marks the85%85\\%recall anchor at which Table[I](https://arxiv.org/html/2609.21061#S5.T1)is read\.![Refer to caption](https://arxiv.org/html/2609.21061v1/Figures/nfa_vs_recall.png)Fig\. 3:False alarms against recall on the blind test split, seed 42, logarithmic ordinate\. Derived from the same precision\-recall data and verified against the tabulated counts at the operating point\.
### V\-BDomain Adaptation is the Dominant Effect

Table[I](https://arxiv.org/html/2609.21061#S5.T1)isolates each component’s contribution, and the largest effect by a wide margin is the domain adaptation itself\. The frozen DINOv3 ViT\-L, probed linearly over unadapted features, reaches0\.300±0\.0110\.300\\pm 0\.011AUPRC, worse than a ResNet18 trained from scratch; injecting LoRA adapters into the same frozen backbone raises this to0\.679±0\.0270\.679\\pm 0\.027, a2\.3×2\.3\\timesimprovement\. This indicates that optical features do not transfer to acoustic imagery: the adaptation, not the backbone, carries the result\.

Comparing against the conventional baselines shows a different picture\. We pair results over the same seeds, since multi\-seed results are not directly comparable to the single\-seed numbers we reported previously\. Paired this way, the adapted ViT\-L leads the strongest baseline, TinyViT\-21M pretrained on ImageNet\-22k with distillation, by\+0\.046±0\.113\+0\.046\\pm 0\.113AUPRC atr=4r=4and\+0\.039±0\.100\+0\.039\\pm 0\.100atr=64r=64; both intervals span zero, although the adapted arm is ahead at all three seeds\. Pretraining behaves the same way, worth\+0\.094±0\.098\+0\.094\\pm 0\.098for TinyViT and\+0\.027±0\.035\+0\.027\\pm 0\.035for ResNet18 against their from\-scratch counterparts, again positive at every seed and again not resolvable\. We therefore claim only what separates:the adapted foundation model is not distinguishable from a well\-initialized conventional CNN at this sample size\. The comparison is not like\-for\-like on pretraining either, with DINOv3 carrying1\.6891\.689B web images against roughly1414M for TinyViT and1\.281\.28M for ResNet18\. That the frozen ViT\-L nonetheless scores0\.3000\.300is the sharper observation: two orders of magnitude more pretraining data buys nothing until it is adapted\.

### V\-CAdapter Capacity Saturates, and Matches Full Fine\-Tuning

We swept the LoRA rank overr∈\{2,4,8,16,32,64\}r\\in\\\{2,4,8,16,32,64\\\}and selectedr=64r=64on validation AUPRC\. On the blind test set performance is*flat*across this range: paired per\-seed,r=4r=4againstr=64r=64gives\+0\.0061±0\.0465\+0\.0061\\pm 0\.0465AUPRC withr=4r=4nominally ahead, and onlyr=2r=2falls away, to0\.62290\.6229\. So786786K trainable parameters,0\.26%0\.26\\%of the backbone, match the12\.5812\.58M ofr=64r=64, an empirical confirmation of the low intrinsic rank hypothesis of Hu et al\.\[[6](https://arxiv.org/html/2609.21061#bib.bib6)\]in the acoustic domain\.

An earlier reading of these experiments, drawn from single seeds, held that low\-rank adaptation matches full fine\-tuning only above some backbone size; replicating every arm at three seeds removes that condition\. On the small architectural variant \(ViT\-S\) \(0\.612±0\.0500\.612\\pm 0\.050atr=64r=64\) a full fine\-tune of all 21\.6M parameters is worth a paired−0\.001±0\.151\-0\.001\\pm 0\.151; at the Large scale \(ViT\-L\), fine\-tuning all 300M gives0\.698±0\.0460\.698\\pm 0\.046against0\.679±0\.0270\.679\\pm 0\.027, a paired\+0\.018±0\.060\+0\.018\\pm 0\.060\. Neither scale separates, and the rank curve is flat at both\. The apparent ViT\-S steepening we previously reported came from a single seed whoser=4r=4draw was the lowest of its three\.

The LoRA\-adapted ViT\-L is stable across experiments:adapting0\.26%0\.26\\%of a ViT\-L, or0\.7%0\.7\\%of a ViT\-S, is not distinguishable from updating every weight, and rank beyondr=4r\{=\}4buys nothing at either size\. One asymmetry emerges: full fine\-tuning is the noisier regime, with a seed standard deviation of0\.0190\.019against0\.0110\.011at ViT\-L and0\.0420\.042against0\.0200\.020at ViT\-S, so equal expected performance at roughly half the variance is the practical case for the adapter\.

### V\-DHard Negative Mining Does Not Separate From Random Selection

Stage 2 mines theK=5,000K=5\{,\}000clutter snippets that the Stage 1 encoder scores as most target\-like and fine\-tunes on those plus all known targets; to test whether mining contributes anything beyond rebalancing exposure, we ran an otherwise identical control drawing its5,0005\{,\}000snippets*at random*\. The two arms do not separate\. Across three seeds, the paired difference between mined and random curricula is−0\.0045±0\.0119\-0\.0045\\pm 0\.0119AUPRC,−0\.0017±0\.0314\-0\.0017\\pm 0\.0314MCC and−12±139\-12\\pm 139false alarms, all spanning zero; the only difference clearing its interval is AUROC, at−0\.0013±0\.0007\-0\.0013\\pm 0\.0007, and it*favours the random control*at all three seeds\. Neither curriculum improves on Stage 1 \(\+0\.0024±0\.0253\+0\.0024\\pm 0\.0253mined,\+0\.0069±0\.0182\+0\.0069\\pm 0\.0182random\)\.

At this curriculum size the intelligence of the mining step is not measurable, and the curriculum’s effect comes from exposure balancing, which random selection supplies equally well\. A plausible mechanism is that mining operates on the training split, where the Stage 1 encoder already achieves99\.9%99\.9\\%AUROC: the snippets that remain difficult under geographic shift are, by construction, not the ones that are difficult on data the model has already fit\.

The obvious objection is that this null result reflects a limitation of the adapter rather than the absence of a mining signal: at12\.5812\.58M trainable parameters, the refiner may simply be unable to express what the curriculum encodes\. We tested this by unfreezing the last four transformer blocks of the otherwise frozen backbone in*both*arms, a five\-fold increase to62\.9762\.97M trainable parameters, while holding the curricula fixed so that capacity is the only thing that varies\. The null does not move: the paired difference is−0\.0045±0\.0061\-0\.0045\\pm 0\.0061AUPRC unfrozen against−0\.0040±0\.0117\-0\.0040\\pm 0\.0117frozen, a difference of differences of−0\.0005±0\.0080\-0\.0005\\pm 0\.0080\. The added capacity buys nothing in either arm on its own either, worth\+0\.0004±0\.0090\+0\.0004\\pm 0\.0090to the mined curriculum and\+0\.0010±0\.0010\+0\.0010\\pm 0\.0010to the random one, which is the rank saturation of Section[V\-C](https://arxiv.org/html/2609.21061#S5.SS3)reached through a different lever\.

### V\-EContrastive Regularization Has Little Left to Enforce

Stage 3 applies a supervised contrastive objective to Stage 2 representations\. Across three seeds it changes little in either direction\. Compared to the baseline runs, paired per\-seed, the model achieves an AUPRC difference of\+0\.0002±0\.0096\+0\.0002\\pm 0\.0096against the mined Stage 2 initialization,−0\.0043±0\.0069\-0\.0043\\pm 0\.0069against the random control, and\+0\.0026±0\.0250\+0\.0026\\pm 0\.0250against Stage 1\. Its MCC∗of0\.643±0\.0150\.643\\pm 0\.015is nominally the highest in Table[I](https://arxiv.org/html/2609.21061#S5.T1), but it overlaps ther=4r\{=\}4arm\.

We previously reported that this stage degrades the high\-recall regime\. Replication does not support that claim: the stage is inert rather than harmful\. Against Stage 2, AUROC changed by only−0\.0033±0\.0127\-0\.0033\\pm 0\.0127andFPR95\\text\{FPR\}\_\{95\}by\+0\.0095±0\.0260\+0\.0095\\pm 0\.0260\. The original claim was prompted by a single low outlier of0\.97950\.9795, the lowest of three seeds, against a replicated mean of0\.98560\.9856\.

One possible explanation lies in the training loss itself\. The loss floor𝔼⁡\[log⁡\|P⁡\(i\)\|\]\\mathbb\{E\}\[\\log\|P\(i\)\|\]is4\.284\.28at our batch size and target fraction, while a structureless embedding scores5\.345\.34under the same projection\. On the encoders it actually receives, Stage 3 begins at a loss between4\.404\.40and4\.624\.62\. That is7575to91%91\\%of the distance from a structureless state to a class\-collapsed one, and the loss barely changes thereafter, so the preceding supervised stages have already imposed most of the geometry the contrastive term exists to enforce\. Because the random control begins at a similar value \(4\.554\.55\), this geometry is a property of supervised adaptation itself rather than of the mined curriculum\.

### V\-FOperating Point, Calibration and Threshold Transfer

The85%85\\%recall anchor is a reporting convention rather than an operational requirement, so Fig\.[2](https://arxiv.org/html/2609.21061#S5.F2)and Fig\.[3](https://arxiv.org/html/2609.21061#S5.F3)give the full trade\-off\. Fig\.[3](https://arxiv.org/html/2609.21061#S5.F3)uses a logarithmic ordinate because the unadapted control’s false\-alarm load runs an order of magnitude above every adapted arm\. The separation gained from domain adaptation holds across the entire recall range, while the individual adapted arms remain visually indistinguishable over most of the curve, the same conclusion the intervals of Table[I](https://arxiv.org/html/2609.21061#S5.T1)reach\. The false\-alarm load rises sharply beyond about0\.90\.9recall, so the marginal cost of the last few points of recall is steep and worth quoting in any operational requirement\.

The later stages contribute one benefit invisible to the ranking metrics: expected calibration error falls from0\.01160\.0116at Stage 1 to0\.00300\.0030after Stage 2 and0\.00360\.0036after Stage 3, and the random control does*not*reproduce this \(0\.01520\.0152\), making calibration the one axis on which mining separates from random selection\. Expected Calibration Error \(ECE\) is computed over1515equal\-width bins on the confidencemax⁡\(p,1−p\)\\max\(p,1\{\-\}p\), with no temperature scaling or isotonic regression at any stage, so these are properties of the trained classifier and not of a post\-hoc correction\.

Threshold transfer is inexpensive in aggregate: the validation\-tuned choice costs11to2020false alarms more than an oracle threshold at the same recall, under4%4\\%of the total in every arm\. Per platform it is not\. Decomposing the blind split by sensor serial across three platforms, carrying244244,135135and5252of the431431targets, the adapted model’s recall spans0\.6920\.692to0\.9060\.906about a global0\.8280\.828, so a deployment requiring85%85\\%recall on*every*platform would not obtain it from an operating point whose aggregate satisfies the requirement\. The spread is not a property of our three\-stage framework: the frozen control ranges0\.7630\.763to0\.9510\.951about0\.8860\.886and TinyViT0\.6350\.635to0\.8730\.873about0\.7730\.773, and the easiest platform is the same under all three architectures, which points at differences between the underlying surveys rather than at any model\. Per\-survey calibration is therefore worth its cost where a recall floor is contractual\. We claim the spread and not a ranking: the ordering of the two weaker platforms reverses between the adapted and frozen arms, and the weakest carries only5252targets\.

### V\-GLimitations of the Protocol

Two properties of the protocol above qualify the comparisons we have drawn\. Neither changes a reported number, and we state both because a reader cannot recover either from the results table\.

The self\-supervised arm sees more imagery than the arms it is compared against\.Contrastive pretraining is label\-free, so we follow standard practice and fit the CC\-SSL encoder over all 148 mission files, validation and test included, whereas every other arm sees the training missions alone\. No annotation from the blind missions reaches any model, but that encoder has seen the unlabeled*imagery*, so its row is transductive and not strictly like\-for\-like\. A second asymmetry runs opposite, since it is scored through a linear probe on frozen features while the supervised arms are optimized end\-to\-end\. The two are indistinguishable in AUPRC, and both asymmetries bear on that comparison in opposite directions\.

The validation split is used for three successive selection decisions\.Adapter rank is chosen on validation AUPRC, and Stages 2 and 3 each select their checkpoint by early stopping on the same split\. Each is individually leakage\-free with respect to the test missions, but they are not independent, they rest on only 362 validation targets, and the resulting optimism is not captured by the test metrics\. The magnitude is not negligible: under a validation\-tuned threshold, test recall drifted from0\.7730\.773to0\.9260\.926across arms, indicating substantial shift between validation and test missions before any selection is stacked on top\. Reading our headline metrics at a recall\-matched operating point insulates the false\-alarm counts from this bias, but it does not protect the choice of rank or checkpoint\. For replications, carve a dedicated selection partition out of the training missions and reserve the validation split solely for establishing the final operating point; rotating geographic folds between stages should be avoided, since doing so would confound stage effects with geographic variance\.

## VIDiscussion

In this work we re\-evaluated a three\-stage parameter\-efficient adaptation framework for SAS ATR, with every stage measured against its own matched control and every close comparison repeated across three seeds\. The performance gains are dominated almost entirely by the initial domain\-adaptation \(LoRA\) stage\. That single step delivers a decisive\+0\.38\+0\.38AUPRC improvement over a frozen backbone, and as little as0\.26%0\.26\\%of the model parameters is enough to obtain it\.

Neither downstream stage separates from its control: hard negative mining is worth−0\.0045±0\.0119\-0\.0045\\pm 0\.0119AUPRC against a random curriculum of equal size, and supervised contrastive refinement\+0\.0002±0\.0096\+0\.0002\\pm 0\.0096against the stage it refines\. We report both null results in full, with the mechanisms established in Sections[V\-D](https://arxiv.org/html/2609.21061#S5.SS4)and[V\-E](https://arxiv.org/html/2609.21061#S5.SS5); each follows from where the stage sits in the pipeline rather than from this dataset\. The one axis on which mining does separate is calibration \(Section[V\-F](https://arxiv.org/html/2609.21061#S5.SS6)\)\.

Our replication also overturns an earlier reading of ours, that parameter efficiency requires backbone scale: no separation is resolved between low\-rank adaptation and full fine\-tuning at either scale we tested, and the rank curve is flat at both\. Taken together, these results carry a concrete engineering implication\. Effort is better spent optimizing the core adaptation step, and the operating point at which the system is evaluated, than on designing elaborate multi\-stage training curricula, since the curriculum stacked on top of a well\-adapted backbone did not repay its cost at this data scale\.

Future work will extend the framework to Cluster\-Level Contrastive Learning\[[10](https://arxiv.org/html/2609.21061#bib.bib20)\]for unsupervised discovery of novel debris types, mine hard negatives from data the encoder has not already fit, and apply the contrastive objective earlier in training\.

## References

- \[1\]\(2002\)SMOTE: synthetic minority over\-sampling technique\.ArXivabs/1106\.1813\.Cited by:[§II](https://arxiv.org/html/2609.21061#S2.p3.1)\.
- \[2\]T\. Chen, S\. Kornblith, M\. Norouzi, and G\. Hinton\(2020\)A simple framework for contrastive learning of visual representations\.ArXivabs/2002\.05709\.Cited by:[§II](https://arxiv.org/html/2609.21061#S2.p3.1)\.
- \[3\]A\. Dosovitskiy, L\. Beyer, A\. Kolesnikov, D\. Weissenborn, X\. Zhai, T\. Unterthiner, M\. Dehghani, M\. Minderer, G\. Heigold, S\. Gelly, J\. Uszkoreit, and N\. Houlsby\(2020\)An image is worth 16x16 words: transformers for image recognition at scale\.ArXivabs/2010\.11929\.Cited by:[§II](https://arxiv.org/html/2609.21061#S2.p2.1)\.
- \[4\]I\. D\. Gerg and V\. Monga\(2021\)Real\-time, deep synthetic aperture sonar \(sas\) autofocus\.In2021 IEEE International Geoscience and Remote Sensing Symposium IGARSS,Vol\.,pp\. 8684–8687\.External Links:[Document](https://dx.doi.org/10.1109/IGARSS47720.2021.9554141)Cited by:[§I](https://arxiv.org/html/2609.21061#S1.p1.1)\.
- \[5\]K\. He, X\. Zhang, S\. Ren, and J\. Sun\(2016\)Deep residual learning for image recognition\.In2016 IEEE Conference on Computer Vision and Pattern Recognition \(CVPR\),Vol\.,pp\. 770–778\.External Links:[Document](https://dx.doi.org/10.1109/CVPR.2016.90)Cited by:[§II](https://arxiv.org/html/2609.21061#S2.p1.1)\.
- \[6\]E\. J\. Hu, Y\. Shen, P\. Wallis, Z\. Allen\-Zhu, Y\. Li, S\. Wang, L\. Wang, and W\. Chen\(2021\)LoRA: low\-rank adaptation of large language models\.arXivabs/2106\.09685\.Cited by:[§I](https://arxiv.org/html/2609.21061#S1.p2.1),[§II](https://arxiv.org/html/2609.21061#S2.p2.1),[§V\-C](https://arxiv.org/html/2609.21061#S5.SS3.p1.1)\.
- \[7\]P\. Khosla, P\. Teterwak, C\. Wang, A\. Sarna, Y\. Tian, P\. Isola, A\. Maschinot, C\. Liu, and D\. Krishnan\(2021\)Supervised contrastive learning\.arXiv2004\.11362\.Cited by:[§I](https://arxiv.org/html/2609.21061#S1.p2.1),[§II](https://arxiv.org/html/2609.21061#S2.p3.1),[§III\-C](https://arxiv.org/html/2609.21061#S3.SS3.p1.1)\.
- \[8\]D\. Kohntopp, B\. Lehmann, D\. Kraus, and A\. Birk\(2017\)Seafloor classification for mine countermeasures operations using synthetic aperture sonar images\.InOCEANS 2017 \- Aberdeen,Vol\.,pp\. 1–5\.External Links:[Document](https://dx.doi.org/10.1109/OCEANSE.2017.8084752)Cited by:[§I](https://arxiv.org/html/2609.21061#S1.p1.1),[§II](https://arxiv.org/html/2609.21061#S2.p1.1)\.
- \[9\]G\. Li, Y\. Gao, X\. Huang, and B\. W\. Ling\(2025\)A hard negatives mining and enhancing method for multi\-modal contrastive learning\.Electronics14\(4\)\.External Links:ISSN 2079\-9292Cited by:[§II](https://arxiv.org/html/2609.21061#S2.p3.1)\.
- \[10\]Y\. Li, P\. Hu, Z\. Liu, D\. Peng, J\. T\. Zhou, and X\. Peng\(2020\)Contrastive clustering\.arXivabs/2009\.09687\.Cited by:[§II](https://arxiv.org/html/2609.21061#S2.p3.1),[§III\-A](https://arxiv.org/html/2609.21061#S3.SS1.p3.1),[§VI](https://arxiv.org/html/2609.21061#S6.p4.1)\.
- \[11\]C\. Lin, L\. Li, D\. Li, J\. Zou, W\. Xue, and Y\. Guo\(2024\)NoRA: nested low\-rank adaptation for efficient fine\-tuning large models\.arXivabs/2408\.10280\.Cited by:[§II](https://arxiv.org/html/2609.21061#S2.p2.1)\.
- \[12\]T\. Lin, P\. Goyal, R\. Girshick, K\. He, and P\. Dollár\(2017\)Focal loss for dense object detection\.In2017 IEEE International Conference on Computer Vision \(ICCV\),Vol\.,pp\. 2999–3007\.Cited by:[§II](https://arxiv.org/html/2609.21061#S2.p3.1)\.
- \[13\]S\. Mangrulkar, S\. Gugger, L\. Debut, Y\. Belkada, S\. Paul, B\. Bossan, and M\. Tietz\(2022\)PEFT: state\-of\-the\-art parameter\-efficient fine\-tuning methods\.Note:https://github\.com/huggingface/peftCited by:[§II](https://arxiv.org/html/2609.21061#S2.p2.1)\.
- \[14\]B\.W\. Matthews\(1975\)Comparison of the predicted and observed secondary structure of t4 phage lysozyme\.Biochimica et Biophysica Acta \(BBA\) \- Protein Structure405\(2\),pp\. 442–451\.Cited by:[§II](https://arxiv.org/html/2609.21061#S2.p3.1)\.
- \[15\]M\. Oquab, T\. Darcet, T\. Moutakanni, H\. Vo, M\. Szafraniec, V\. Khalidov, P\. Fernandez, D\. Haziza, F\. Massa, A\. El\-Nouby, M\. Assran, N\. Ballas, W\. Galuba, R\. Howes, P\. Huang, S\. Li, I\. Misra, M\. Rabbat, V\. Sharma, G\. Synnaeve, H\. Xu, H\. Jegou, J\. Mairal, P\. Labatut, A\. Joulin, and P\. Bojanowski\(2024\)DINOv2: learning robust visual features without supervision\.arXivabs/2304\.07193\.Cited by:[§II](https://arxiv.org/html/2609.21061#S2.p2.1)\.
- \[16\]S\. Rajbhandari, J\. Rasley, O\. Ruwase, and Y\. He\(2020\)ZeRO: memory optimizations toward training trillion parameter models\.arXivabs/1910\.02054\.Cited by:[§II](https://arxiv.org/html/2609.21061#S2.p2.1)\.
- \[17\]J\. Rasley, S\. Rajbhandari, O\. Ruwase, and Y\. He\(2020\)DeepSpeed: system optimizations enable training deep learning models with over 100 billion parameters\.Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining\.Cited by:[§II](https://arxiv.org/html/2609.21061#S2.p2.1)\.
- \[18\]T\. Ridnik, E\. Ben\-Baruch, N\. Zamir, A\. Noy, I\. Friedman, M\. Protter, and L\. Zelnik\-Manor\(2021\)Asymmetric loss for multi\-label classification\.InProceedings of the IEEE/CVF International Conference on Computer Vision \(ICCV\),Note:arXiv:2009\.14119Cited by:[§III\-A](https://arxiv.org/html/2609.21061#S3.SS1.p2.1),[§III\-B](https://arxiv.org/html/2609.21061#S3.SS2.p1.1),[§IV\-C](https://arxiv.org/html/2609.21061#S4.SS3.p2.1)\.
- \[19\]O\. Russakovsky, J\. Deng, H\. Su, J\. Krause, S\. Satheesh, S\. Ma, Z\. Huang, A\. Karpathy, A\. Khosla, M\. Bernstein, A\. C\. Berg, and L\. Fei\-Fei\(2015\)ImageNet Large Scale Visual Recognition Challenge\.International Journal of Computer Vision \(IJCV\)115\(3\),pp\. 211–252\.External Links:[Document](https://dx.doi.org/10.1007/s11263-015-0816-y)Cited by:[§II](https://arxiv.org/html/2609.21061#S2.p2.1)\.
- \[20\]O\. Siméoni, H\. V\. Vo, M\. Seitzer, F\. Baldassarre, M\. Oquab, C\. Jose, V\. Khalidov, M\. Szafraniec, S\. Yi, M\. Ramamonjisoa, F\. Massa, D\. Haziza, L\. Wehrstedt, J\. Wang, T\. Darcet, T\. Moutakanni, L\. Sentana, C\. Roberts, A\. Vedaldi, J\. Tolan, J\. Brandt, C\. Couprie, J\. Mairal, H\. Jégou, P\. Labatut, and P\. Bojanowski\(2025\)DINOv3\.arXivabs/2508\.10104\.Cited by:[§I](https://arxiv.org/html/2609.21061#S1.p2.1),[§II](https://arxiv.org/html/2609.21061#S2.p2.1)\.
- \[21\]I\. J\. Sledge, M\. S\. Emigh, J\. L\. King, D\. L\. Woods, J\. T\. Cobb, and J\. C\. Príncipe\(2022\)Target detection and segmentation in circular\-scan synthetic aperture sonar images using semisupervised convolutional encoder–decoders\.IEEE Journal of Oceanic Engineering47\(4\),pp\. 1099–1128\.External Links:[Document](https://dx.doi.org/10.1109/JOE.2022.3152863)Cited by:[§I](https://arxiv.org/html/2609.21061#S1.p1.1),[§II](https://arxiv.org/html/2609.21061#S2.p1.1)\.
- \[22\]M\. Valipour, M\. Rezagholizadeh, I\. Kobyzev, and A\. Ghodsi\(2023\)DyLoRA: parameter\-efficient tuning of pre\-trained models using dynamic search\-free low\-rank adaptation\.InProceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics \(EACL\),pp\. 3274–3287\.Note:arXiv:2210\.07558Cited by:[§II](https://arxiv.org/html/2609.21061#S2.p2.1)\.
- \[23\]G\. D\. Vetaw, F\. E\. Bobe, and D\. Bryner\(2026\)Contrastive learning for synthetic aperture sonar pattern recognition with vision transformers\.Proceedings of the Institution of Acoustics\.Cited by:[§II](https://arxiv.org/html/2609.21061#S2.p3.1)\.
- \[24\]G\. D\. Vetaw, B\. Cowen, D\. C\. Brown, D\. P\. Williams, and S\. Jayasuriya\(2025\)Volumetric hadamard normalization for synthetic aperture sonar sub\-bottom target classification\.IEEE Journal of Oceanic Engineering50\(4\),pp\. 3024–3038\.Cited by:[§I](https://arxiv.org/html/2609.21061#S1.p1.1),[§II](https://arxiv.org/html/2609.21061#S2.p1.1)\.
- \[25\]D\. P\. Williams and D\. C\. Brown\(2025\)Spectral partitioning of synthetic aperture sonar imagery for improved atr\.IEEE Geoscience and Remote Sensing Letters\(\),pp\. 1–1\.External Links:[Document](https://dx.doi.org/10.1109/LGRS.2025.3554335)Cited by:[§I](https://arxiv.org/html/2609.21061#S1.p1.1),[§II](https://arxiv.org/html/2609.21061#S2.p1.1)\.
- \[26\]D\. P\. Williams\(2016\)Underwater target classification in synthetic aperture sonar imagery using deep convolutional neural networks\.In23rd International Conference on Pattern Recognition \(ICPR\),Vol\.,pp\. 2497–2502\.External Links:[Document](https://dx.doi.org/10.1109/ICPR.2016.7900011)Cited by:[§I](https://arxiv.org/html/2609.21061#S1.p1.1),[§II](https://arxiv.org/html/2609.21061#S2.p1.1)\.
- \[27\]N\. Yang, G\. Li, S\. Wang, Z\. Wei, H\. Ren, X\. Zhang, and Y\. Pei\(2025\)SS\-yolo: a lightweight deep learning model focused on side\-scan sonar target detection\.Journal of Marine Science and Engineering13\(1\)\.External Links:ISSN 2077\-1312Cited by:[§II](https://arxiv.org/html/2609.21061#S2.p1.1)\.
- \[28\]S\. Yang, X\. Zhang, and P\. Tan\(2026\)FS2\-detr: transformer\-based few\-shot sonar object detection with enhanced feature perception\.Journal of Marine Science and Engineering\.Cited by:[§I](https://arxiv.org/html/2609.21061#S1.p1.1),[§II](https://arxiv.org/html/2609.21061#S2.p1.1)\.
- \[29\]J\. Zhang and B\. Gao\(2025\)RCDI\-yolo: a target\-detection method for complex environment side\-scan sonar images based on improved yolov8\.Frontiers in Marine Science12\.Cited by:[§I](https://arxiv.org/html/2609.21061#S1.p1.1),[§II](https://arxiv.org/html/2609.21061#S2.p1.1)\.