As one of the most important underwater sensing technologies, forward-looking sonar exhibits unique imaging characteristics. Sonar images are often affected by severe speckle noise, low texture contrast, acoustic shadows, and geometric distortions. These factors make it difficult for traditional teacher-student frameworks to achieve satisfactory performance in sonar semantic segmentation tasks under extremely limited labeled data conditions. To address this issue, we propose a Collaborative Teacher Semantic Segmentation Framework for forward-looking sonar images. This framework introduces a multi-teacher collaborative mechanism composed of one general teacher and multiple sonar-specific teachers. By adopting a multi-teacher alternating guidance strategy, the student model can learn general semantic representations while simultaneously capturing the unique characteristics of sonar images, thereby achieving more comprehensive and robust feature modeling. Considering the challenges of sonar images, which can lead teachers to generate a large number of noisy pseudo-labels, we further design a cross-teacher reliability assessment mechanism. This mechanism dynamically quantifies the reliability of pseudo-labels by evaluating the consistency and stability of predictions across multiple views and multiple teachers, thereby mitigating the negative impact caused by noisy pseudo-labels. Notably, on the FLSMD dataset, when only 2% of the data is labeled, our method achieves a 5.08% improvement in mIoU compared to other state-of-the-art approaches.
Figures & tables
Figure 1: The unique characteristics of sonar images lead to limited adaptability in conventional teacher–student frameworks, making them ineffective for handling semantic segmentation tasks under extremely scarce annotation conditions.
Figure 2: (a) The overall architecture of CTFS, where knowledge is transferred to the student through the collaboration between the traditional teacher and the sonar teacher. (b) The multi-view reliability assessment process of pseudo-labels.
Figure 3: (a) Example of shadows in sonar images: due to the obstruction of objects during sonar propagation, a shadow is formed behind the object. The green hollow area represents the GT, and the red area represents the shadow behind the target. (b) Example of how sonar energy decreases with propagation distance. In the right-hand image, the solid blue line represents the vertical energy distribution in the left-hand image, and the dashed red line is a linear regression fit to the blue line.
Figure 4: (a) The collection process of the FSSG dataset. (b) Sample distribution of each category in the FSSG dataset, and the visualization of targets under sonar and visible perspectives.
Figure 5: Qualitative demonstrations of different approaches on the FLSMD dataset with 2% labeled data.
Figure 6: Qualitative demonstrations of different approaches on the FSSG dataset with 2% labeled data.
Method
Venue
Encoder
FLSMD
FSSG
2%
5%
10%
2%
5%
10%
Labeled Only
–
RN-101
46.47
55.15
61.94
17.64
32.23
52.27
Labeled Only
–
DINOv2-S
51.08
61.02
68.48
35.61
41.67
54.58
AEL [ 14 ]
NIPS’21
RN-101
52.70
64.89
70.73
53.51
57.84
63.08
Dual Teacher [ 24 ]
NIPS’23
MiT-B2
56.23
66.09
69.61
58.31
60.16
65.91
UniMatch V1 [ 38 ]
CVPR’23
RN-101
50.70
56.14
61.66
58.39
60.87
63.15
Table 1: Quantitative comparison of mIoU (%) on the FLSMD and FSSG datasets. Experiments on both the FLSMD and FSSG datasets were conducted with labeled data of 2%, 5%, and 10%. The best results are shown in bold , and the second best results are shown in italic .
Figure 7: Performance comparison of tail-class categories on the FLSMD dataset with a 2% labeled and the FSSG dataset with a 5% labeled. (a) performance comparison on the hook category, (b) performance comparison on the shampoo-bottle category, (c) performance comparison on the tire category, (d) performance comparison on the underwater robot category.
Model
EMA
CBTS
MVRA
mIoU
M1
-
-
-
51.08
M2
√
-
-
55.94
M3
√
√
-
59.94
M4
√
√
√
62.32
Table 2: Ablation study on key components of CTFS conducted on the FLSMD dataset with 2% labeled data.
#
Method
mIoU
Hook
Shampoo Bottle
1
Tgeneral
55.94
60.94
55.07
2
#1 + Tsonar_a
57.12
62.83
70.56
3
#2 + Tsonar_b
59.94
65.59
74.40
Table 3: Ablation study on the effectiveness of the sonar-specific perturbations designed for the two additional sonar teachers on the FLSMD dataset with 2% labeled data.
Figure 8: mIoU results for each parameter on the FLSMD dataset with 5% labeled data.
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.
Taqi Hamoda, Hayat Rajani, Nuno Gracias
Computer Vision and Robotics Research Institute (ViCOROB) of the University of Girona, Spain.
Supervised deep learning achieves strong performance in biomedical image segmentation but relies on costly pixel-wise annotations, motivating semi-supervised approaches that exploit unlabeled data. We introduce a diffusion-based teacher--student framework in which segmentation predictions are used to condition image denoising, encouraging the production of more informative pseudo-labels. The teacher is first pretrained through an unsupervised reconstruction task using diffusion-style corruption, timestep conditioning, and denoising. Starting from a corrupted empty mask, the model predicts an intermediate segmentation that conditions image denoising, encouraging the predicted mask to capture structural information useful for recovering the original image. The resulting teacher is then co-trained with a student using supervised segmentation on labeled samples and cross pseudo-supervision on unlabeled data. We further introduce a multi-round extension during co-training, in which the teacher generates multiple stochastic image reconstructions and corresponding segmentation predictions, providing additional reconstruction and alignment signals to improve its pseudo-labels. We evaluate the proposed framework on three public 2D biomedical segmentation datasets and a 3D left atrial segmentation benchmark. Across several labeling regimes, our method achieves competitive or superior performance compared with state-of-the-art semi-supervised approaches, with the largest gains observed under severe label scarcity.
Sonar imaging is the primary modality for underwater target detection, yet small targets remain difficult to detect due to insufficient pixel coverage, low acoustic contrast, and scale ambiguity across imaging ranges. CNN-based detectors extract local features efficiently but cannot suppress noise-induced false alarms without global acoustic context. Transformer-based methods capture long-range dependencies at quadratic computational cost. Existing Mamba-based vision models offer efficient linear-cost scanning but lack multi-scale semantic alignment across pyramid levels, multi-receptive-field fusion, and small-target-aware training supervision needed for reliable sonar detection. This letter proposes Mamba Dilated-Scale Fusion (MambaDSF), a hybrid framework addressing these limitations through three contributions: a Mamba Enhanced Feature Pyramid (MambaEFP) backbone that jointly captures local echo cues and global acoustic context at linear complexity; a Dilate Fusion Mamba (DFMamba) encoder that enforces multi-scale feature alignment across pyramid levels; and Scale-Adaptive Weighted IoU (SA-WIoU) and Cross-Scale Coherence (CSC) losses that stabilize small-target training. MambaDSF achieves 91.5% mAP50 on the UATD forward-looking sonar benchmark with 28.7 million parameters, surpassing all compared detectors. On a small-target subset the gain reached +2.2 percentage points, and cross-domain evaluation on FLS and MD-FLS confirms the generalization of the proposed architecture. The codes are publicly available at https://github.com/IDontKnowAAA/MambaDSF.
Hui Lin, Jiayi Li, Jing Wang +1
School of Information Science and Engineering, Ocean University of China, Qingdao 266100, China · School of Information and Communication Technology, Griffith University, Nathan, QLD 4111, Australia