TERA: Task-Evidence Restoration for Underwater Object Detection under Bandwidth-Constrained Transmission

Heejin Kim; Joon-Seok Song; Eunil Park

cs.CVNov 2, 2026

TERA: Task-Evidence Restoration for Underwater Object Detection under Bandwidth-Constrained Transmission

Authors: Heejin Kim, Joon-Seok Song, Eunil Park

Organizations: Sungkyunkwan University, Seoul, Republic of Korea · Jaume I University, Castellon, Spain

Abstract

Underwater object detectors are typically evaluated on pre-transmission images, overlooking additional degradation from bandwidth-constrained downscaling and compression. We introduce AquaTX, a controlled benchmark built from real underwater detection datasets with seen and held-out codecs. We also propose TERA, a metadata-blind receiver-side framework combining task-guided pixel restoration and dual-state feature correction while keeping the detector frozen. Under seen compound degradation, TERA improves mean average precision (mAP50:95, averaged over IoU thresholds from 0.50 to 0.95) from 37.7 to 50.0 on DUO-TX and from 31.5 to 41.3 on RUOD-TX relative to the frozen detector without restoration. TERA also preserves clean-input performance and remains competitive with detector fine-tuning under the held-out AVIF codec. These findings support transmission-aware evaluation and task-guided restoration for underwater object detection. Code is available at https://github.com/hijinkim/TERA.

Figures & tables

Figure 1: Detection under bandwidth-constrained transmission. (a) Pre-transmission image with ground truth. (b) Transmitted image after downscaling and compression. (c) Missed detections by the frozen detector. (d) TERA restores task evidence and recovers the missed detections without updating the detector.

Dataset	Split	Images	Boxes
DUO-TX	Train	6,003	57,574
	Val	668	6,424
	Test	1,111	10,517
RUOD-TX	Train	2,712	21,619
	Val	301	2,275
	Test	1,255	10,553

Table 1: Statistics of the underwater source datasets used in AquaTX. DUO-TX uses the original four-class DUO dataset, while RUOD-TX uses the four-class RUOD-4C subset. Both are evaluated independently under the same transmission protocol.

Condition group	Factors	Validation	Test
Clean reference	-	1	1
Resize-only	3 sizes	3	3
Codec-only	codecs $\times$ 3 rates	9	12
Compound	codecs $\times$ 3 sizes $\times$ 3 rates	27	36
Total		40	52

Table 2: AquaTX evaluation conditions. Validation uses the three seen codecs, whereas test evaluation additionally includes the unseen AVIF codec.

Figure 2: Overview of TERA. The pixel restorer uses intermediate backbone $P_{3}/P_{4}$ features from the degraded path for task-guided RGB restoration, whereas the evidence restorer operates on later neck features $N_{3}/N_{4}$ . The degraded and restored images are forwarded through the same frozen backbone and neck to obtain two late feature states. A dual-state evidence restorer jointly exploits these states to refine the $N_{3}$ and $N_{4}$ pyramid levels, while $N_{5}$ follows the restored-image path unchanged. Only the pixel and evidence restorers are trainable.

DUO-TX
Strategy	Method	Clean	Resize-only	Codec-only	Seen comp.	Unseen AVIF	$\Delta$ Clean
		(1)	(3)	(9)	(27)	(9)
No adaptation	Clean-YOLO	65.7	60.3	57.0	37.7	42.2	0.0
Image restoration	UWCNN [ 21 ] (FT)	63.4	59.6	56.0	38.3	43.0	-2.3
	NU2Net [ 15 ] (FT)	64.2	60.4	56.5	38.9	43.4	-1.5
	EDTR (FT) [ 19 ]	33.6	33.2	32.4	27.3	28.7	-32.1

Table 3: Main detection results on AquaTX in mAP 50:95 (%). Clean-YOLO is trained on the original pre-transmission images and directly evaluated without restoration or degradation adaptation. Image-restoration methods operate in front of the same frozen Clean-YOLO detector, whereas DegFT-YOLO adapts the detector itself. Resize-only, codec-only, and seen-compound results are macro-averaged over 3, 9, and 27 conditions, respectively. Unseen AVIF denotes the nine compound conditions generated using the held-out AVIF codec; the three AVIF codec-only controls are excluded from the main aggregate.

\Delta

Clean is the clean-input change relative to Clean-YOLO. Best and second-best results are shown in bold and underline, respectively. Zero-shot results and additional enhancement baselines are reported in the supplementary material.

Figure 3: Incremental gain of full TERA over pixel-only restoration under varying transmission severity. When averaged over the remaining transmission factors, the gain over pixel-only restoration is larger at lower coding rates (left) and smaller transmitted rasters (right) on both datasets.

Figure 4: Qualitative ablation of TERA. Columns (a)–(d) show the clean reference with ground truth, the transmitted input, the result after pixel-only restoration, and the final detection result obtained using full TERA, respectively. The first row visualizes detector response maps before box decoding, where warmer colors indicate stronger object evidence. The second row shows the corresponding detection results. Compared with pixel-only restoration, full TERA strengthens object responses and recovers additional detections.

Configuration	Seen	Clean	Unseen
	Comp.		AVIF
Raw	37.7	65.7	42.2
Pixel only	48.3	66.0	51.6
+ Dual-state restoration	49.6	66.6	52.8
+ Foreground KD (TERA)	50.0	66.4	53.2
Feature-L1 objective	45.9	65.5	49.9

Table 4: Ablation of TERA on DUO-TX in mAP 50:95 (%). Seen Comp. and Unseen AVIF are macro-averaged over 27 and 9 conditions, respectively.

Underwater Object Detection

Underwater object detection localizes and classifies targets in degraded camera and sonar data for marine science, robotics, and inspection.

See all 30 papers

A Dual-Branch Collaborative Framework for Joint Optimization of Underwater Image Enhancement and Object Detection
Jun 14, 2026Liyuan Cao, Zheng Liu, Guanghao Liao +2Underwater Image Enhancement Underwater Imaging
School of Electronic and Information Engineering, University of Science and Technology Liaoning, No.189 Qianshan Middle Road, Lishan District, Anshan, 114051, Liaoning, China
Adaptive Enhancement and Dual-Pooling Sequential Attention for Lightweight Underwater Object Detection with YOLOv10
Mar 4, 2026Md. Mushibur Rahman, Umme Fawzia Rahim, Enam Ahmed TaufikObject Detection Underwater Object Detection
Department of Computer Science and Engineering, Dhaka University of Engineering & Technology Gazipur, 1707, Bangladesh · Department of Computer Science and Engineering, BRAC University Dhaka, Bangladesh
Degradation-Guided Underwater Image Restoration with Task-Oriented Latent Control
Aug 9, 2026Xu Zhang, Xuhui Cao, Kangzhe Yuan +5Underwater Image Enhancement Underwater Imaging
National Engineering Research Center for Multimedia Software, School of Computer Science, Wuhan University · School of Computer and Artificial Intelligence, Zhengzhou University · Aerospace Information Research Institute, Henan Academy of Sciences +2

Jun 14, 2026J/K move · Enter open · S save

A Dual-Branch Collaborative Framework for Joint Optimization of Underwater Image Enhancement and Object Detection

Liyuan Cao, Zheng Liu, Guanghao Liao, Yonghui Yang+1

School of Electronic and Information Engineering, University of Science and Technology Liaoning, No.189 Qianshan Middle Road, Lishan District, Anshan, 114051, Liaoning, China

Underwater Image Enhancement Underwater Imaging Underwater Object Detection

Open paper AskPDF arXiv

Due to wavelength dependent light absorption and scattering, underwater images usually suffer from color distortion and blurred details, which limits underwater object detection performance. Existing underwater image enhancement methods mainly focus on visual quality improvement, while it is still difficult to balance enhancement quality, processing efficiency, and downstream detection performance. Therefore, this paper proposes an efficient dual-branch underwater image enhancement framework for object detection. The detail enhancement branch improves brightness and local contrast to recover texture details in dark regions. The color restoration branch uses adaptive compensation to reduce color distortion and improve color gradation. By combining the complementary outputs of the two branches, the proposed framework provides clearer and more informative images for object detection. On the UIEB and EUVP datasets, the proposed method achieves UIQM scores of 2.249 and 2.576. When applied to the YOLOv8 detection task on the URPC dataset, the proposed method improves mAP50 by 2.1% compared with the baseline. Extensive experiments show that our method improves object detection in complex underwater scenes, while balancing enhancement quality and processing efficiency.

Abstract

Underwater Object Detection

Explore similar work

A Dual-Branch Collaborative Framework for Joint Optimization of Underwater Image Enhancement and Object Detection

Adaptive Enhancement and Dual-Pooling Sequential Attention for Lightweight Underwater Object Detection with YOLOv10

Degradation-Guided Underwater Image Restoration with Task-Oriented Latent Control