Mine-JEPA: In-Domain Self-Supervised Learning for Mine-Like Object Classification in Side-Scan Sonar

Taeyoun Kwon; Youngwon Choi; Hyeonyu Kim; Myeongkyun Cho; Junhyeok Choi; Moon Hwan Kim

cs.CVApr 1, 2026

Mine-JEPA: In-Domain Self-Supervised Learning for Mine-Like Object Classification in Side-Scan Sonar

Authors: Taeyoun Kwon, Youngwon Choi, Hyeonyu Kim, Myeongkyun Cho, Junhyeok Choi, Moon Hwan Kim

Organizations: Maum AI Inc. · Seoul National University · KAIST · Yonsei University

Abstract

Side-scan sonar (SSS) mine classification is a challenging maritime vision problem characterized by extreme data scarcity and a large domain gap from natural images. While self-supervised learning (SSL) and general-purpose vision foundation models have shown strong performance in general vision and several specialized domains, their use in SSS remains largely unexplored. We present Mine-JEPA, the first in-domain SSL pipeline for SSS mine classification, using SIGReg, a regularization-based SSL loss, to pretrain on only 1,170 unlabeled sonar images. In the binary mine vs. non-mine setting, Mine-JEPA achieves an F1 score of 0.935, outperforming fine-tuned DINOv3 (0.922), a foundation model pretrained on 1.7B images. For 3-class mine-like object classification, Mine-JEPA reaches 0.820 with synthetic data augmentation, again outperforming fine-tuned DINOv3 (0.810). We further observe that applying in-domain SSL to foundation models degrades performance by 10--13 percentage points, suggesting that stronger pretrained models do not always benefit from additional domain adaptation. In addition, Mine-JEPA with a compact ViT-Tiny backbone achieves competitive performance while using 4x fewer parameters than DINOv3. These results suggest that carefully designed in-domain self-supervised learning is a viable alternative to much larger foundation models in data-scarce maritime sonar imagery.

Figures & tables

Figure 1: Performance comparison across model scales. Using only 1,170 sonar images for pretraining, Mine-JEPA outperforms DINOv3, pretrained on 1.7B images, at both backbone scales. ViT-Tiny remains competitive with $4\times$ fewer parameters.

Figure 2: Overview of the Mine-JEPA pipeline. Stage 1 extracts unlabeled $96\times 96$ patches from full SSS images using a sliding window. Stage 2 performs in-domain self-supervised pretraining with SSS-adapted multi-view augmentation and the SIGReg objective. Stage 3 evaluates the pretrained backbone with frozen or fine-tuned probes for downstream patch classification.

Figure 3: Representative samples from the public side-scan sonar dataset. Top: a full SSS image with annotated mine-like contacts (MILCO) and non-mine bottom objects (NOMBO). Bottom: $96\times 96$ patches from the three classes used in our evaluation: MILCO, background (BG), and NOMBO.

Method	Init	SSL Data	Best Probe	macro-F1	NOMBO F1	Acc
Random Init	Random	—	ft_mlp	0.557 $\pm$ 0.052	0.439 $\pm$ 0.091	58.3 $\pm$ 4.0%
IN1K Only	IN1K	—	ft_mlp	0.739 $\pm$ 0.028	0.647 $\pm$ 0.046	76.3 $\pm$ 2.6%
DINOv3	DINOv3	—	ft_mlp	0.810 $\pm$ 0.025	0.700 $\pm$ 0.068	83.4 $\pm$ 2.0%
SimCLR	IN1K	Real+Syn	finetune	0.801 $\pm$ 0.038	0.693 $\pm$ 0.084	82.6 $\pm$ 2.6%
VICReg	IN1K	Real+Syn	finetune	0.800 $\pm$ 0.037	0.676 $\pm$ 0.068	82.8 $\pm$ 3.0%
BYOL	IN1K	Real+Syn	finetune	0.693 $\pm$ 0.028	0.539 $\pm$ 0.059	72.8 $\pm$ 2.4%

Table 1: 3-class patch classification on BG/MILCO/NOMBO. For each method, we report the best probe mode and the corresponding macro-F1, NOMBO F1, and accuracy. Probe modes: linear and mlp (frozen backbone), finetune and ft mlp (fine-tuned backbone).

Hyperparameter	Value
Projection dim.	16
$\lambda$	0.1
Views ( $V$ )	4
Batch size	1024
Optimizer	AdamW (weight decay 0.05)
Learning rate	$1.4\times 10^{-3}$

Table 2: Default SSL pretraining configuration (ViT-S/16). All architectures use the same setup unless noted otherwise.

Method	Init	Params	3-class F1	2-class F1	MILCO Recall (%)
DINOv3	DINOv3	21.5M	0.810 $\pm$ 0.025	0.922 $\pm$ 0.018	88.1 $\pm$ 1.3
Mine-JEPA (ViT-S)	IN1K	21.6M	0.820 $\pm$ 0.018	0.935 $\pm$ 0.020	90.9 $\pm$ 2.4
Mine-JEPA (ViT-Tiny)	IN1K	5.5M	0.814 $\pm$ 0.026	0.935 $\pm$ 0.020	91.4 $\pm$ 2.3

Table 3: Model comparison across 3-class and binary settings. Mine-JEPA uses Real+Syn data for 3-class classification and Real data for binary classification. For each method, we report the best probe mode.

Config	Init	SSL Data	$\boldsymbol{\lambda}$	macro-F1	$\boldsymbol{\Delta}$
Random Init	Random	—	—	0.557 $\pm$ 0.052	—
Natural Aug SSL	Random	Real ∗	0.02	0.312 $\pm$ 0.040	$-24.5$ %p
DINOv3 + SIGReg	DINOv3	Real	0.02	0.706 $\pm$ 0.068	$-10.4$ %p vs DINOv3
DINOv3 + SIGReg	DINOv3	Real+Syn	0.02	0.677 $\pm$ 0.031	$-13.3$ %p vs DINOv3
SIGReg (SSS aug)	Random	Real	0.02	0.725 $\pm$ 0.021	—
+ IN1K init	IN1K	Real	0.02	0.756 $\pm$ 0.054	+3.1%p vs Random SSL

Table 4: Ablation on initialization, augmentation, and data composition. Real ∗ denotes natural-image augmentation (saturation/hue jitter, solarization, grayscale). Rows 1–2 compare against Random Init, rows 3–4 against DINOv3, and rows 5–8 show cumulative gains; each

\Delta

is relative to the preceding row.

SSL Method	Loss Type	proj. dim	Real only	Real+Syn	$\boldsymbol{\Delta}$
SIGReg (Ours)	Regularization	16	0.799 $\pm$ 0.036	0.820 $\pm$ 0.018	+2.1%p
VICReg	Regularization	2048	0.774 $\pm$ 0.025	0.800 $\pm$ 0.037	+2.6%p
SimCLR	Contrastive	128	0.806 $\pm$ 0.013	0.801 $\pm$ 0.038	$-0.5$ %p
BYOL	Distillation	256	0.774 $\pm$ 0.031	0.693 $\pm$ 0.028	$-8.1$ %p

Table 5: SSL method comparison and data scalability. All methods use ViT-S/16 with IN1K initialization, 100 epochs. macro-F1 on 3-class classification.

\Delta

= Real+Syn

-

Real only.

Classification

Classification maps observations to discrete labels, including familiar, unknown, hierarchical, or multiple categories.

See all 408 papers

BenthicDINO: Physics-Informed Self-Distillation for View-Invariant Side-Scan Sonar Representations
Aug 24, 2026Taqi Hamoda, Hayat Rajani, Nuno GraciasRepresentation Learning Invariant Representation Learning
Computer Vision and Robotics Research Institute (ViCOROB) of the University of Girona, Spain.
Adapting Vision Foundation Models to Acoustics for Pose-Free 3D Sonar Reconstruction
Sep 5, 2026Kevin Zhang, Jingxi Chen, Mohamad Qadri +5Underwater 3D Reconstruction 3D Reconstruction
University of Maryland, College Park · Carnegie Mellon University · University of Rhode Island +1
LoRA Enhanced Contrastive Learning with SAS Vision Transformers
Sep 17, 2026Dan Zimmerman, Frank E. Bobe, Amelia L. McCormack +2SAR ATR Underwater Object Detection
Center for Connected Autonomy & AI Florida Atlantic University Boca Raton, USA · Naval Surface Warfare Center Panama City Division Panama City, USA · Department of Computer Science Florida State University Tallahassee, USA

Aug 24, 2026J/K move · Enter open · S save

BenthicDINO: Physics-Informed Self-Distillation for View-Invariant Side-Scan Sonar Representations

Taqi Hamoda, Hayat Rajani, Nuno Gracias

Computer Vision and Robotics Research Institute (ViCOROB) of the University of Girona, Spain.

Representation Learning Invariant Representation Learning Self-Supervised Visual Representation Learning Physics-Informed ML

Open paper AskPDF arXiv

Automated interpretation of side-scan sonar (SSS) imagery is traditionally hindered by view-dependence, as recorded intensities inextricably conflate intrinsic seabed reflectivity with transient acquisition geometries and acoustic artifacts. To address this, we introduce BenthicDINO, a physics-informed self-supervised learning framework that decouples acquisition dynamics from seabed representations. Utilizing a ConvNeXt-v2 backbone, physics-guided augmentations, and an HSIC penalty, BenthicDINO explicitly enforces feature invariance to physical viewing parameters like slant range and incidence angle. Evaluations demonstrate exceptional data efficiency - achieving near-peak performance in benthic habitat mapping with only 10% of available labels - and strong cross-domain generalization, outperforming fully supervised baselines on the AI4Shipwrecks benchmark using frozen representations.

Abstract

Classification

Explore similar work

BenthicDINO: Physics-Informed Self-Distillation for View-Invariant Side-Scan Sonar Representations

Adapting Vision Foundation Models to Acoustics for Pose-Free 3D Sonar Reconstruction

LoRA Enhanced Contrastive Learning with SAS Vision Transformers