Downstream Task Inspired Underwater Image Enhancement: A Perception-Aware Study from Dataset Construction to Network Design

Bosen Lin; Feng Gao; Yanwei Yu; Junyu Dong; Qian Du

cs.CVMar 2, 2026

Downstream Task Inspired Underwater Image Enhancement: A Perception-Aware Study from Dataset Construction to Network Design

Authors: Bosen Lin, Feng Gao, Yanwei Yu, Junyu Dong, Qian Du

Organizations: State Key Laboratory of Physical Oceanography, Ocean University of China, Qingdao 266100, China. · Department of Electrical and Computer Engineering, Mississippi State University, Starkville, MS 39762 USA.

Abstract

In real underwater environments, downstream image recognition tasks such as semantic segmentation and object detection often face challenges posed by problems like blurring and color inconsistencies. Underwater image enhancement (UIE) has emerged as a promising preprocessing approach, aiming to improve the recognizability of targets in underwater images. However, most existing UIE methods mainly focus on enhancing images for human visual perception, frequently failing to reconstruct high-frequency details that are critical for task-specific recognition. To address this issue, we propose a Downstream Task-Inspired Underwater Image Enhancement (DTI-UIE) framework, which leverages human visual perception model to enhance images effectively for underwater vision tasks. Specifically, we design an efficient two-branch network with task-aware attention module for feature mixing. The network benefits from a multi-stage training framework and a task-driven perceptual loss. Additionally, inspired by human perception, we automatically construct a Task-Inspired UIE Dataset (TI-UIED) using various task-specific networks. Experimental results demonstrate that DTI-UIE significantly improves task performance by generating preprocessed images that are beneficial for downstream tasks such as semantic segmentation, object detection, and instance segmentation. The codes are publicly available at https://github.com/oucailab/DTIUIE.

Figures & tables

Fig. 1: Framework of the proposed DTI-UIE. (a) A task-inspired UIE dataset is constructed automatically by task networks. (b) A UIE network providing advantages for downstream tasks is achieved with task-relevant priors and task-driven perceptual loss.

Fig. 2: Underwater images enhanced using different UIE methods. UColor [ 16 ] and HUPE [ 17 ] blurred the outlines of the diver and background, and introduced additional edge and texture information to the foreground seabed. The proposed DTI-UIE introduced more accurate foreground-background differentiation and reduced texture and edge noise, which is beneficial for downstream tasks.

Dataset	Year	Img. Num.	GT & Raw Generation Methods	Degrade & Enhance Method	Objective
Fabbri et al. [ 27 ]	2018	7941	Degradation from HQ images	CycleGAN	Perception-centered
EUVP [ 29 ]	2020	11435	Degradation from HQ images	CycleGAN	Perception-centered
UFO-120 [ 28 ]	2020	1620	Degradation from HQ images	CycleGAN + GB	Perception-centered
UIEB [ 3 ]	2020	890	Majority-voting from Enc by human	12 Trad methods	Perception-centered
SUIM-E [ 18 ]	2022	1635	Majority-voting from Enc by human	12 Trad methods	Perception-centered
HICRD [ 30 ]	2022	2000	Imaging parameter inversion	Trad method	Perception-centered

TABLE I: Comparison of different UIE benchmark datasets. The degrade or enhance methods (Trad) refer to traditional underwater image enhancement methods, (DL) refers to deep-learning-based underwater image enhancement methods, (GB) refers to Gaussian blurring. The image source (Enc) refer to enhanced images, and (HQ) refer to high-quality underwater images

Fig. 3: (a) The percentage of images from different UIE sources corresponding to ground truth in the entire TI-UIED dataset. (b) Number of images containing each object category.

Fig. 4: Example raw and references images in the TI-UIED dataset.

Fig. 5: Overview of the proposed DTI-UIE framework, which consists of the enhancement network $\mathbf{Enc}$ and two task network $\mathbf{Seg}_{pri}$ , $\mathbf{Seg}_{task}$ . In the first stage, the task network $\mathbf{Seg}_{pri}$ for task-relevant prior is trained. During the second stage, DTI-UIE updates the enhancement network $\mathbf{Enc}$ using the TDP loss and pixel loss, while both two task networks are frozen. In the third stage, the task network $\mathbf{Seg}_{task}$ for TDP loss is updated, using different kinds of enhanced and mixed images.

Fig. 6: Structure of TA-CTB. At the $i$ -th layer, the TA-CAM mixs the target-relevant prior $T_{i}$ and image feature map $X_{i}$ and produces the output feature map $\hat{X}_{i}$ .

Fig. 7: Visual comparison with different UIE methods on SUIM [ 13 ] (1-3) and LIACi [ 7 ] (4-6) dataset.

Fig. 8: Segmentation results on SUIM [ 13 ] dataset. The results are obtained by training the UNet [ 44 ] on raw and different enhanced images.

Fig. 9: Segmentation results on LIACi [ 7 ] dataset. The results are obtained by training the UNet [ 44 ] on raw and different enhanced images.

Method	SUIM Dataset [ 13 ]								LIACi Dataset [ 7 ]
	UNet [ 44 ]		DeepLabv3 [ 47 ]		FPN [ 50 ]		DPT [ 60 ]		UNet [ 44 ]		DeepLabv3 [ 47 ]		FPN [ 50 ]		DPT [ 60 ]
	Dice	mIoU	Dice	mIoU	Dice	mIoU	Dice	mIoU	Dice	mIoU	Dice	mIoU	Dice	mIoU	Dice	mIoU
Raw	72.29	66.52	78.02	66.51	74.96	64.74	82.48	73.18	75.55	54.43	82.51	55.35	81.63	53.66	81.11	56.88
UDCP [ 58 ]	69.58	62.12	77.45	62.28	73.84	60.70	82.08	68.69	76.80	52.21	80.64	53.87	80.44	50.35	81.89	53.19
IBLA [ 59 ]	72.09	61.67	78.10	61.81	76.27	62.71	81.22	68.91	76.43	49.18	79.75	50.25	78.74	45.89	80.12	50.94
CBLA [ 22 ]	72.16	63.63	75.04	67.02	75.07	62.79	81.82	70.73	77.85	54.15	82.67	55.22	79.81	50.95	79.00	55.30

TABLE II: Comparison for semantic segmentation in terms of mIoU(

\uparrow

) and Dice(

\uparrow

) on SUIM [ 13 ] and LIACi [ 7 ] dataset. The best and second result are marked in bold and underline .

Fig. 10: Visualization results of object detection on RUOD dataset [ 64 ] . The results are obtained by training the SSD [ 62 ] network on the raw image and different enhanced images.

Method	Faster-RCNN [ 61 ]						SSD [ 62 ]						DINO [ 63 ]
Method	mAPs	mAPm	mAPl	mAP50	mAP75	mAP	mAPs	mAPm	mAPl	mAP50	mAP75	mAP	mAPs	mAPm	mAPl	mAP50	mAP75	mAP
Raw	14.10	40.70	57.30	81.60	57.40	52.10	15.10	35.80	55.90	80.80	54.30	50.50	19.30	43.70	64.80	83.60	63.50	59.00
UDCP [ 58 ]	16.80	36.90	53.80	79.30	52.70	48.80	11.90	34.10	54.50	79.60	52.60	49.30	19.00	42.70	63.70	83.10	62.10	58.00
IBLA [ 59 ]	14.40	35.60	52.40	77.90	50.50	47.40	14.00	33.10	53.30	78.60	51.00	48.20	18.00	39.20	60.30	79.70	58.20	54.60
CBLA [ 22 ]	13.80	37.90	56.30	80.40	55.40	50.80	12.60	30.90	51.10	77.10	48.20	46.00	19.40	40.50	62.60	60.60	81.50	56.60
NU2Net [ 25 ]	16.00	39.70	58.00	82.00	57.60	52.40	8.20	34.00	53.80	79.80	50.90	48.50	18.40	43.00	64.40	83.10	62.90	58.30

TABLE III: Comparison for object detection in terms of mAPs(

\uparrow

), mAPm(

\uparrow

) mAPl(

\uparrow

), mAP50(

\uparrow

), mAP75(

\uparrow

) and mAP(

\uparrow

) on RUOD [ 64 ] dataset. The best and second result are marked in bold and underline .

Fig. 11: Visualization results of instance segmentation on UIIS dataset [ 37 ] . The results are obtained by training the Mask R-CNN [ 61 ] network on the raw image and different enhanced images.

Method	Mask R-CNN [ 61 ]						Water Mask R-CNN [ 37 ]
Method	mAPs	mAPm	mAPl	mAP50	mAP75	mAP	mAPs	mAPm	mAPl	mAP50	mAP75	mAP
Raw	7.50	17.70	29.70	40.10	20.50	21.90	8.30	19.00	34.40	40.00	23.90	23.30
UDCP [ 58 ]	7.70	18.60	28.20	40.10	19.80	20.30	6.70	17.10	29.60	36.50	19.50	20.00
IBLA [ 59 ]	7.20	18.60	28.50	38.70	20.60	20.60	6.30	17.20	29.60	35.90	20.70	20.10
CBLA [ 22 ]	6.70	17.70	28.30	38.70	19.90	20.60	7.10	18.60	32.20	38.50	22.40	21.90
NU2Net [ 25 ]	7.60	19.00	29.60	39.50	22.00	21.60	7.70	18.90	32.60	39.20	23.70	22.60

TABLE IV: Comparison for instance segmentation in terms of mAP50(

\uparrow

), mAP75(

\uparrow

) and mAP(

\uparrow

) on UIIS [ 37 ] dataset. The best and second result are marked in bold and underline .

Method	SUIM-E [ 18 ]			UIEB [ 3 ]
Method	UIQM	UCIQE	NUIQ	UIQM	UCIQE	NUIQ
UDCP [ 58 ]	1.81	62.17	6.34	2.14	59.49	6.06
IBLA [ 59 ]	1.87	62.49	6.72	2.34	53.81	6.34
CBLA [ 22 ]	2.24	67.44	6.75	2.32	68.72	7.45
NU2Net [ 25 ]	3.00	58.41	5.74	3.34	58.07	5.99
WfDiff [ 26 ]	2.66	58.35	6.45	3.25	53.26	6.57

TABLE V: Comparison for image quality assessment in terms of UIQM [ 66 ] (

\uparrow

), UCIQE [ 66 ] (

\uparrow

) and NUIQ [ 67 ] (

\uparrow

) on SUIM-E [ 18 ] and UIEB [ 3 ] . The best and second result are marked in bold and underline .

Method	Params (M)	FLOPs (G)
NU2Net [ 25 ]	3.1	10.4
WfDiff [ 26 ]	100.5	369.2
SemiUIR	7.1	36.45
CCMSR [ 23 ]	21.6	43.3
HUPE [ 17 ]	12.1	162.2
TFUIE [ 10 ]	1.3	25.26

TABLE VI: Comparison for different deep-learning-based UIE methods in terms of model size and computational complexity.

Methods	Dice	mIoU
DTI-UIE without FRB branch	72.89	67.01
DTI-UIE without DEB branch	72.52	66.41
DTI-UIE ( Ours )	74.48	67.61

TABLE VII: Effectiveness of the FRB and DEB. The comparison by the implementation of different network branches.

Methods	Dice	mIoU
DTI-UIE without TA-CTB	71.15	67.13
DTI-UIE ( Ours )	74.48	67.61

TABLE VIII: Effectiveness of the TA-CTB. The comparison by including or excluding TA-CTB and target-relevant prior.

Methods	Dice	mIoU
DTI-UIE without TDP loss	71.71	65.12
DTI-UIE with no trainable TDP loss	73.27	66.50
DTI-UIE ( Ours )	74.48	67.61

TABLE IX: Effectiveness of the TDP loss. The comparison by including or excluding TDP loss, as well as by not updating TDP loss.

Methods	Dice	mIoU
Mixing with $16\times 16$ block	72.96	65.94
Mixing with $8\times 8$ block ( Ours )	74.48	67.61
Mixing with $1\times 1$ pixel	73.90	65.53
Full image mixing	73.50	65.32
No mixing in TDP loss	71.34	64.82

TABLE X: Effectiveness the mixing process in TDP loss. The comparison by varying the granularity of image mixing on TDP loss.

Methods	Dice	mIoU
Merging Task Networks	69.58	65.59
Separating Task Networks( Ours )	74.48	67.61

TABLE XI: Effectiveness of Task Network Number. The comparison by merging or separating

\mathbf{Seg}_{pri}

and

\mathbf{Seg}_{task}

network.

Methods	Dice	mIoU
ResNet50 UNet	74.58	66.44
VGG-13 UNet	71.64	66.90
VGG-16 UNet ( Ours )	74.48	67.61

TABLE XII: Effectiveness of Task Network Structure. The comparison by the structure of

\mathbf{Seg}_{pri}

and

\mathbf{Seg}_{task}

network.

Methods	Dice	mIoU
DTI-UIE trained on SUIM-E	70.47	65.18
DTI-UIE trained on TI-UIED ( Ours )	74.48	67.61
WfDiff source pretrained	72.37	66.09
WfDiff trained on TI-UIED	73.61	66.25
CCMSR source pretrained	72.58	65.47
CCMSR trained on TI-UIED	70.50	63.96

TABLE XIII: Effectiveness of the TI-UIED. The comparison by training on TI-UIED dataset or SUIM-E dataset.

Explore similar work

Empowering Semantic-Sensitive Underwater Image Enhancement with VLM
Mar 13, 2026Guodong Fan, Shengning Zhou, Genji Yuan +3Image Enhancement Underwater Image Enhancement
Shandong Technology and Business University, Yantai, China · Shandong University of Finance and Economics, Jinan, China · Dalian Maritime University, Dalian, China
Style-Decoupled Adaptive Routing Network for Underwater Image Enhancement
Apr 14, 2026Hang Xu, Chen Long, Bing Wang +2Image Enhancement Underwater Image Enhancement
State Key Laboratory of Information Engineering in Surveying, Mapping and Remote Sensing, Wuhan University, Wuhan, 430079, China · Department of Aeronautical and Aviation Engineering, The Hong Kong Polytechnic University, Hung Hom, Kowloon, Hong Kong Special Administrative Region of China
Hero-Mamba: Mamba-based Dual Domain Learning for Underwater Image Enhancement
Apr 17, 2026Tejeswar Pokuri, Shivarth RaiUnderwater Image Enhancement Underwater Imaging
Manipal Institute of Technology, Manipal Academy of Higher Education, Manipal - 576104, India

Mar 13, 2026J/K move · Enter open · S save

Empowering Semantic-Sensitive Underwater Image Enhancement with VLM

Guodong Fan, Shengning Zhou, Genji Yuan, Huiyu Li+2

Shandong Technology and Business University, Yantai, China · Shandong University of Finance and Economics, Jinan, China · Dalian Maritime University, Dalian, China

Image Enhancement Underwater Image Enhancement Underwater Imaging Image Restoration

Open paper AskPDF arXiv

In recent years, learning-based underwater image enhancement (UIE) techniques have rapidly evolved. However, distribution shifts between high-quality enhanced outputs and natural images can hinder semantic cue extraction for downstream vision tasks, thereby limiting the adaptability of existing enhancement models. To address this challenge, this work proposes a new learning mechanism that leverages Vision-Language Models (VLMs) to empower UIE models with semantic-sensitive capabilities. To be concrete, our strategy first generates textual descriptions of key objects from a degraded image via VLMs. Subsequently, a text-image alignment model remaps these relevant descriptions back onto the image to produce a spatial semantic guidance map. This map then steers the UIE network through a dual-guidance mechanism, which combines cross-attention and an explicit alignment loss. This forces the network to focus its restorative power on semantic-sensitive regions during image reconstruction, rather than pursuing a globally uniform improvement, thereby ensuring the faithful restoration of key object features. Experiments confirm that when our strategy is applied to different UIE baselines, significantly boosts their performance on perceptual quality metrics as well as enhances their performance on detection and segmentation tasks, validating its effectiveness and adaptability.

Abstract

Explore similar work

Empowering Semantic-Sensitive Underwater Image Enhancement with VLM

Style-Decoupled Adaptive Routing Network for Underwater Image Enhancement

Hero-Mamba: Mamba-based Dual Domain Learning for Underwater Image Enhancement