Task-Driven Underwater Image Enhancement

• The paper introduces DTI-UIE, a task-oriented UIE framework that leverages a task-inspired dataset and dual-branch network to optimize downstream vision tasks.
• The framework employs a dual-branch enhancement network with task-aware attention and a task-driven perceptual loss to preserve recognition-relevant features.
• Building on TI-UIED, the method demonstrates consistent gains in semantic segmentation, object detection, and instance segmentation compared to conventional UIE methods.

Searching arXiv for the requested topic and key related papers to ground the article in current literature. arxiv_search query: "Downstream Task Inspired Underwater Image Enhancement (Lin et al., 2 Mar 2026)" Downstream Task-Inspired Underwater Image Enhancement (DTI-UIE) is a task-oriented underwater image enhancement framework that treats enhancement as a preprocessing problem for machine perception rather than primarily as a problem of human visual pleasantness. In this formulation, underwater enhancement is designed to preserve and reconstruct features that matter to semantic segmentation, object detection, and instance segmentation, rather than only color realism, global contrast, or aesthetically pleasing sharpness. The framework couples a task-inspired dataset, a dual-branch enhancement network, task-prior injection, and a task-driven perceptual loss, and reports consistent gains across multiple downstream underwater vision tasks (Lin et al., 2 Mar 2026).

1. Conceptual framing and research context

DTI-UIE is motivated by an empirical mismatch: many underwater image enhancement methods improve color, contrast, and subjective visibility, but these gains do not reliably translate into better downstream recognition performance, and can even degrade it. The underlying claim is that human-oriented UIE frequently fails to reconstruct the high-frequency details that are critical for task-specific recognition, including object boundaries, local texture cues, edge consistency, foreground-background separability, and semantically stable structure. The framework therefore redefines enhancement quality in terms of recognizability for underwater vision tasks rather than visual pleasantness alone (Lin et al., 2 Mar 2026).

This framing places DTI-UIE within a broader shift in underwater enhancement research. Some semantically guided methods already move beyond pure pixel regression, but do so indirectly. For example, CLIP-guided UIE uses a CLIP perception model and curriculum contrastive regularization to align enhancement with semantic and perceptual structure, yet it is described as “not a direct task-driven UIE method” because it does not optimize detection or segmentation during training (Cao et al., 29 Apr 2025). Similarly, a dual-branch collaborative enhancement framework for object detection is explicitly detection-oriented and reports improved YOLOv8 performance, but it is characterized as “detection-oriented enhancement design with task-based model selection, rather than true joint enhancement-detection optimization through shared gradients” (Cao et al., 14 Jun 2026). DTI-UIE is more explicit in its task alignment: dataset construction, network priors, and perceptual supervision are all tied to downstream task behavior (Lin et al., 2 Mar 2026).

A central misconception addressed by this line of work is that better-looking underwater images are necessarily better for machine vision. DTI-UIE reports that conventional human-visual-quality methods may blur object contours, alter textures in misleading ways, or emphasize visually pleasing but semantically irrelevant content, whereas task-oriented enhancement should preserve recognition-friendly boundaries and stable feature representations (Lin et al., 2 Mar 2026). Related studies reinforce the same discrepancy: detection-oriented enhancement reports that “improved visual quality does not always lead to better downstream object detection performance” (Cao et al., 14 Jun 2026), and physically decomposed, transferred-prior enhancement argues that downstream visual algorithms require clear structure and correct color rather than only human-preferred appearance (Hu et al., 14 Jun 2026).

2. TI-UIED: task-inspired dataset construction

A core contribution of DTI-UIE is TI-UIED, described as the first UIE dataset explicitly tailored for downstream vision tasks (Lin et al., 2 Mar 2026). Instead of selecting references by human voting, the dataset replaces human preference with task-network preference. In earlier UIE datasets, multiple enhanced candidates are shown to human annotators and the preferred image is chosen as the reference. TI-UIED preserves the same procedural logic but changes the evaluator: semantic segmentation networks, rather than human observers, determine which candidate enhancement is most useful.

The source images come from SUIM, which contains 1635 real underwater images with pixel-level annotations for eight semantic categories: water body, fish, reefs, aquatic plants, wrecks/ruins, human divers, robots, and sea-floor. For training the task-specific evaluators used during dataset construction, UIIS is used as the training set and SUIM as the target set (Lin et al., 2 Mar 2026).

For each raw image, the candidate pool contains ten inputs in total: five traditional UIE methods, four deep UIE methods, and the raw image itself. The traditional methods are CLAHE, Fusion, ULAP, WWPE/WWPF, and WB. The deep methods are SGUIE, TOPAL, TUDA, and UColor. Including the raw image is methodologically important, because it allows the construction process to retain the input unchanged if no enhancement improves the task (Lin et al., 2 Mar 2026).

The evaluator ensemble consists of seven semantic segmentation architectures: UNet, SegFormer, PSPNet, DeepLabV3, DeepLabV3+, MaNet, and FPN. Each training process is repeated three times to reduce randomness, yielding $(5+4+1)\times7\times3$ trained models in total across enhancement conditions, architectures, and repetitions. Candidate ranking is based on mIoU, and the final reference is selected by a majority voting mechanism across networks and runs (Lin et al., 2 Mar 2026).

This construction protocol makes TI-UIED task-driven at the reference level. The selected target is not the image that looks best to humans, but the image that most consistently improves segmentation. The paper later reports that DTI-UIE trained on TI-UIED significantly outperforms the same framework trained on SUIM-E, with Dice improving from 70.47 to 74.48 and mIoU from 65.18 to 67.61, indicating that task-inspired reference construction is not an incidental detail but a substantial part of the method’s effect (Lin et al., 2 Mar 2026).

3. Network design: dual-branch enhancement with task-aware attention

DTI-UIE is defined by a task-prior pipeline and a dual-branch enhancement architecture. Given a raw image $I_{raw} \in \mathbb{R}^{H\times W\times 3}$, an enhancement ground truth $I_{gt} \in \mathbb{R}^{H\times W\times 3}$, and a segmentation mask $M_{gt} \in \mathbb{R}^{H\times W\times 1}$, the prior task network produces

$$\hat{T}_{pri} = \mathbf{Seg}_{pri}(I_{raw};\theta_{pri}),$$

the enhancement network produces

$$\hat{I}_{enc} = \mathbf{Enc}(I_{raw},\hat{T}_{pri};\theta_{enc}),$$

and the downstream task prediction is written as

$$\hat{M}_{enc} = \mathbf{Seg}_{task}(\hat{I}_{enc};\theta_{task}).$$

The paper writes the joint objective as

$$\underset{\theta_{enc},\theta_{pri},\theta_{task}}{\min}\ \mathcal{L}_{enc}(\hat{I}_{enc},I_{gt})+\mathcal{L}_{task}(\hat{M}_{enc},M_{gt}).$$

These equations define the basic relation between priors, enhancement, and downstream prediction (Lin et al., 2 Mar 2026).

The enhancement network $\mathbf{Enc}$ has two branches. The Feature Restoration Branch (FRB) is a UNet-shaped encoder-decoder built around Conv-attention Transformer Blocks (CTBs) from Co-Scale Conv-Attentional Image Transformers. A $3\times3$ convolution projects the input into a 32-channel feature map $I_{raw} \in \mathbb{R}^{H\times W\times 3}$0. The encoder contains three hierarchical stages with channel sizes $I_{raw} \in \mathbb{R}^{H\times W\times 3}$1 at resolutions $I_{raw} \in \mathbb{R}^{H\times W\times 3}$2, followed by a bottleneck with 256-channel features at $I_{raw} \in \mathbb{R}^{H\times W\times 3}$3. Downsampling uses a $I_{raw} \in \mathbb{R}^{H\times W\times 3}$4 convolution with stride 2. The decoder uses three Task-Aware Conv-attention Transformer Blocks (TA-CTBs) followed by one CTB, transposed convolution for $I_{raw} \in \mathbb{R}^{H\times W\times 3}$5 upsampling, skip connections via Skip Connection Modules (SCM), a final CTB refinement block, and a $I_{raw} \in \mathbb{R}^{H\times W\times 3}$6 convolution with a global residual connection (Lin et al., 2 Mar 2026).

The Detail Enhancement Branch (DEB) operates at full input resolution and consists of four residual convolution blocks. Each block includes a channel attention module and a pixel attention module. The stated purpose is to preserve textures, edges, and local structures that may be weakened by encoder-decoder compression (Lin et al., 2 Mar 2026).

The principal architectural novelty is the TA-CTB, which injects task-relevant priors from a segmentation network into enhancement features. The task network is a UNet with a VGG-16 backbone, and priors are taken from decoder features. At stage $I_{raw} \in \mathbb{R}^{H\times W\times 3}$7, image features $I_{raw} \in \mathbb{R}^{H\times W\times 3}$8 and task priors $I_{raw} \in \mathbb{R}^{H\times W\times 3}$9 are dimension-aligned, patch-embedded into tokens, and positionally encoded by depthwise convolution. Image tokens generate $I_{gt} \in \mathbb{R}^{H\times W\times 3}$0, $I_{gt} \in \mathbb{R}^{H\times W\times 3}$1, and $I_{gt} \in \mathbb{R}^{H\times W\times 3}$2; task tokens generate $I_{gt} \in \mathbb{R}^{H\times W\times 3}$3; and the mixed task-aware query is

$I_{gt} \in \mathbb{R}^{H\times W\times 3}$4

The block then performs task-aware attention and adds convolutional relative position encoding before a feed-forward update (Lin et al., 2 Mar 2026). The paper’s stated interpretation is that TA-CTB emphasizes regions important for recognition and preserves structures useful to segmentation.

This two-branch split has a direct DTI interpretation. FRB targets global, semantically meaningful, task-relevant representations, whereas DEB preserves the fine local structures that dense and instance-level recognition depend on. The ablation study supports that complementarity: removing FRB reduces Dice to 72.89 and mIoU to 67.01, removing DEB reduces Dice to 72.52 and mIoU to 66.41, and the full model reaches Dice 74.48 and mIoU 67.61 (Lin et al., 2 Mar 2026).

4. Training framework and task-driven perceptual supervision

DTI-UIE uses a three-stage alternate training framework. In Stage 1, the prior task network $I_{gt} \in \mathbb{R}^{H\times W\times 3}$5 is trained on raw underwater images with segmentation cross-entropy: $I_{gt} \in \mathbb{R}^{H\times W\times 3}$6 Its role is to learn stable, target-relevant semantic priors from original underwater images (Lin et al., 2 Mar 2026).

In Stage 2, the enhancement network $I_{gt} \in \mathbb{R}^{H\times W\times 3}$7 is trained while $I_{gt} \in \mathbb{R}^{H\times W\times 3}$8 and $I_{gt} \in \mathbb{R}^{H\times W\times 3}$9 are frozen. The enhancement loss combines a pixel loss and a Task-Driven Perceptual (TDP) loss. The pixel loss is

$M_{gt} \in \mathbb{R}^{H\times W\times 1}$0

and the TDP loss is

$M_{gt} \in \mathbb{R}^{H\times W\times 1}$1

where $M_{gt} \in \mathbb{R}^{H\times W\times 1}$2 denotes the feature extractor of the downstream task network, specifically at the last encoder (Lin et al., 2 Mar 2026).

The TDP loss is a defining aspect of the framework. Standard perceptual losses use classification backbones trained on natural-image categorization, typically VGG-style networks, whereas TDP uses a segmentation network trained on underwater semantic labels. The objective is therefore aligned with underwater dense-prediction structure rather than generic natural-image similarity (Lin et al., 2 Mar 2026). This is what makes the loss “task-driven” in the paper’s terminology.

In Stage 3, the task-loss network $M_{gt} \in \mathbb{R}^{H\times W\times 1}$3 is updated using three image domains: the enhanced image $M_{gt} \in \mathbb{R}^{H\times W\times 1}$4, the reference image $M_{gt} \in \mathbb{R}^{H\times W\times 1}$5, and a mixed image $M_{gt} \in \mathbb{R}^{H\times W\times 1}$6. The mixed image is generated by patchwise blending of $M_{gt} \in \mathbb{R}^{H\times W\times 1}$7 and $M_{gt} \in \mathbb{R}^{H\times W\times 1}$8 using a random mask that is passed through sigmoid and resized by bilinear interpolation. The task-training loss is

$M_{gt} \in \mathbb{R}^{H\times W\times 1}$9

The paper presents this stage as necessary to prevent shortcut learning and to keep task-space supervision reliable as the enhanced-image domain changes (Lin et al., 2 Mar 2026).

Several ablations identify the training schedule as structurally important. Removing TDP reduces Dice from 74.48 to 71.71 and mIoU from 67.61 to 65.12. Using a frozen, non-trainable TDP improves over no TDP but remains below the full model at Dice 73.27 and mIoU 66.50. Merging the two task networks degrades Dice from 74.48 to 69.58 and mIoU from 67.61 to 65.59. The default $$\hat{T}_{pri} = \mathbf{Seg}_{pri}(I_{raw};\theta_{pri}),$$0 patch mixing also outperforms $$\hat{T}_{pri} = \mathbf{Seg}_{pri}(I_{raw};\theta_{pri}),$$1, $$\hat{T}_{pri} = \mathbf{Seg}_{pri}(I_{raw};\theta_{pri}),$$2, full-image mixing, and no mixing (Lin et al., 2 Mar 2026). These results indicate that DTI-UIE is not simply a dual-branch enhancer with an auxiliary segmentation prior; it is a staged co-training framework in which the stability of priors and the adaptability of task-space supervision are treated as separate optimization problems.

5. Empirical performance on enhancement and downstream tasks

DTI-UIE is evaluated on semantic segmentation, object detection, instance segmentation, and UIE image-quality benchmarks. The downstream evaluation spans multiple architectures: UNet, DeepLabV3, FPN, and DPT for semantic segmentation; Faster R-CNN, SSD, and DINO for object detection; and Mask R-CNN and Water Mask R-CNN for instance segmentation (Lin et al., 2 Mar 2026).

On SUIM semantic segmentation, DTI-UIE reports Dice/mIoU of 74.48/67.61 with UNet, 78.93/67.61 with DeepLabV3, 76.85/66.22 with FPN, and 80.64/73.55 with DPT. Compared with raw images, mIoU improves from 66.52 to 67.61 for UNet, from 66.51 to 67.61 for DeepLabV3, from 64.74 to 66.22 for FPN, and from 73.18 to 73.55 for DPT. On LIACi, DTI-UIE reports 78.02/55.27 with UNet, 83.13/55.47 with DeepLabV3, 81.68/53.72 with FPN, and 82.82/57.17 with DPT (Lin et al., 2 Mar 2026).

On RUOD object detection, DTI-UIE improves Faster R-CNN to $$\hat{T}_{pri} = \mathbf{Seg}_{pri}(I_{raw};\theta_{pri}),$$3, $$\hat{T}_{pri} = \mathbf{Seg}_{pri}(I_{raw};\theta_{pri}),$$4, and $$\hat{T}_{pri} = \mathbf{Seg}_{pri}(I_{raw};\theta_{pri}),$$5, compared with raw images at 52.10 mAP. SSD reaches $$\hat{T}_{pri} = \mathbf{Seg}_{pri}(I_{raw};\theta_{pri}),$$6, better than raw at 50.50. DINO reaches $$\hat{T}_{pri} = \mathbf{Seg}_{pri}(I_{raw};\theta_{pri}),$$7, above raw at 59.00 (Lin et al., 2 Mar 2026). The gains are modest, but they are consistent across detector families, which suggests that the enhancement is not overfit to a single task backbone.

On UIIS instance segmentation, DTI-UIE improves Mask R-CNN from 21.90 to 22.30 mAP and Water Mask R-CNN from 23.30 to 23.60 mAP (Lin et al., 2 Mar 2026). The paper interprets the improvements in $$\hat{T}_{pri} = \mathbf{Seg}_{pri}(I_{raw};\theta_{pri}),$$8 and large-object AP as evidence that better structural preservation improves mask quality.

The same model is only competitive, rather than dominant, on human-oriented no-reference IQA. On SUIM-E, DTI-UIE reports UIQM 2.79, UCIQE 54.18, and NUIQ 5.94. On UIEB, it reports UIQM 3.33, UCIQE 54.92, and NUIQ 5.98. Other methods outperform it on some or all of these metrics (Lin et al., 2 Mar 2026). This divergence is one of the paper’s main empirical claims: better human-perception scores do not necessarily imply better downstream task performance.

The ablations further clarify which components matter most. Removing TA-CTB reduces Dice from 74.48 to 71.15 and mIoU from 67.61 to 67.13. Training the same model on SUIM-E instead of TI-UIED reduces Dice from 74.48 to 70.47 and mIoU from 67.61 to 65.18. Retraining other UIE models on TI-UIED gives mixed outcomes, with WfDiff improving slightly while CCMSR degrades (Lin et al., 2 Mar 2026). This suggests that TI-UIED alone is not sufficient; the architecture and the loss design must also be task-aligned.

In computational terms, DTI-UIE reports 9.5M parameters and 361.0G FLOPs. The paper characterizes the model as moderate in parameter count but relatively heavy in computation. Runtime or FPS is not reported (Lin et al., 2 Mar 2026).

6. Position within DTI-UIE research, limitations, and implications

Within the underwater enhancement literature, DTI-UIE is notable because it integrates task orientation at three levels simultaneously: reference construction, network priors, and perceptual supervision. Other recent methods touch one or two of these dimensions but not all. CLIP-guided UIE introduces pretrained semantic guidance but remains “downstream-compatible but not downstream-optimal” because it does not optimize task performance directly (Cao et al., 29 Apr 2025). Detection-oriented enhancement can use object detection to select design parameters, yet still remain an indirect preprocessing method without shared-gradient co-optimization (Cao et al., 14 Jun 2026). Physics- and prior-fusion UIE can be explicitly validated on feature detection, matching, edge detection, and saliency detection, but is still characterized as a “strong supporting paper” rather than a direct DTI-UIE method because it lacks downstream losses (Hu et al., 14 Jun 2026). DTI-UIE differs in that task behavior is already embedded in the chosen references and in the TDP feature space (Lin et al., 2 Mar 2026).

The paper also has clear limitations. TI-UIED references are selected using semantic segmentation, so the dataset construction is task-specific and may inherit segmentation bias. The candidate pool is restricted to outputs of existing enhancement methods plus the raw image, which means the optimal task-friendly reference may not exist among the candidates. Different downstream tasks may prefer different enhancement candidates in borderline cases. The authors therefore identify multi-task dataset construction—jointly using segmentation, detection, and instance segmentation in future reference selection—as a natural extension (Lin et al., 2 Mar 2026).

A second limitation is conceptual. DTI-UIE is task-inspired and task-evaluated, but its internal task machinery is still centered on segmentation. The fact that segmentation-selected references transfer to detection and instance segmentation is empirically important, but it does not imply that a segmentation-driven construction is universally optimal for all underwater tasks. This suggests a broader implication: task-oriented enhancement may require task-family-specific priors and evaluation criteria rather than a single universal definition of “good enhancement.”

The paper’s most consequential conclusion is that underwater enhancement should be treated as a perception-aligned preprocessing problem rather than only an image-restoration problem. In this view, the central quantities are not just UIQM, UCIQE, or NUIQ, but whether enhancement preserves recognition-friendly boundaries, stable feature representations, and machine-usable structures. That claim is consistent with adjacent literature showing that perceptual metrics and downstream performance can diverge (Cao et al., 14 Jun 2026). DTI-UIE formalizes that divergence into a concrete method: construct references by task voting, inject task priors into enhancement, and supervise restoration in a task feature space (Lin et al., 2 Mar 2026).