PaperlayerView exact revision on arXiv

Source-linked AI summary

PIC-UIE: Predicting Image-Adaptive Corrections for Lightweight Underwater Image Enhancement

Cunhao Zhu, Dongliang Xu, Xiangtao Kong, Xiaoyan Lu, Tianyu Wang, Yue Yao

arXiv:2609.33318v1cs.CV

TL;DR

Underwater image enhancement must restore visibility, color fidelity, and structure while remaining practical for high-resolution edge deployment. PIC-UIE predicts structured corrections from a fixed RGB thumbnail and applies them in YCbCr, reporting 3.197–3.257 dB higher UIEB PSNR than matched-capacity dense RGB and YCbCr residual prediction.

  • Problem

    UIE must restore visibility, color fidelity, and structure, while edge deployment requires high-resolution processing within tight compute, memory, and latency budgets.

    Source: Introduction · p. 1

  • Method

    PIC-UIE predicts image-adaptive corrections from a fixed 256 × 256 RGB thumbnail and applies them to the native-resolution input in YCbCr.

    Source: Introduction · p. 1

  • Results

    3.197–3.257 dB higher UIEB PSNR than matched-capacity dense RGB and YCbCr residual prediction, reaching 24.137 dB on UIEB-90 and 21.320 dB on zero-shot LSUI.

    Source: Conclusion · p. 7

  • Takeaways & Limitations

    The results support fixed-size structured prediction as an approach that scales across desktop and edge hardware.

    Source: Conclusion · p. 7

Abstract

from arXiv · show

Underwater image enhancement (UIE) aims to restore visibility, color fidelity, and structural detail from images degraded by wavelength-dependent attenuation and backscatter. State-of-the-art UIE methods often rely on large backbones and dense image-to-image prediction, limiting their practicality for edge deployment. Moreover, operating entirely in a single color space couples degradation estimation with luminance and chroma correction. To address these challenges, we propose PIC-UIE, a lightweight predictor--executor framework that predicts image-adaptive corrections from a fixed $256\times256$ RGB thumbnail and applies them to the native-resolution input in the YCbCr color space. The predictor produces seven outputs, organized into spatial correction, nonlinear luminance and coupled chroma mapping, and image-level color calibration. A depth map regularizes the transmission proxy during training, whereas inference uses only the RGB input. With 9,486 parameters and 0.094 GFLOPs at $256\times256$, PIC-UIE achieves 24.137 dB PSNR and 0.9216 SSIM on UIEB-90 and 21.320 dB PSNR on zero-shot LSUI. It further processes native 4K images at 55.0 FPS under the comparison protocol. These results show that structured correction prediction provides an effective and practical alternative to dense RGB reconstruction for underwater image enhancement.

Introduction

PIC-UIE predicts compact, image-adaptive corrections from a fixed RGB thumbnail and applies them to the native-resolution input in YCbCr. Its structured correction design targets practical underwater enhancement while preserving scene detail and limiting prediction cost.

Introduction

  • Wavelength-dependent attenuation and backscatter cause color casts, veiling haze, and contrast loss, motivating enhancement that restores visibility and color fidelity while preserving scene structure.The degradation varies with scene distance and water conditions (Akkaynak and Treibitz 2018, 2019); embedded cameras and robotic platforms also face tight compute, memory, and latency budgets.Source: Introduction · p. 1
  • 9,486 parameters and 0.094 GFLOPs at 256 × 256 accompany 24.137 dB PSNR and 0.9216 SSIM on UIEB-90, 21.320 dB PSNR on zero-shot LSUI, and 55.0 FPS on native 4K.The quality–efficiency trade-off is summarized in Figure 1.Source: Figure 1 · p. 1, Introduction · p. 2
  • Native-resolution deployment is costly for methods using large backbones, multi-stage processing, or dense prediction, while RGB prediction and YCbCr correction exploit distinct representation strengths.RGB captures channel-wise attenuation cues, whereas YCbCr separates luminance from chroma.Source: Introduction · p. 1
  • PIC-UIE predicts seven outputs from a 256 × 256 RGB thumbnail and executes them on the native-resolution image in YCbCr, with depth regularization used only during training.The outputs cover spatial correction, nonlinear luminance–chroma mapping, and image-level color calibration; inference requires RGB alone.Source: Introduction · pp. 1–2

Related Work

Prior UIE work explores learning-based restoration, complementary color representations, efficient image-adaptive operators, and physical or geometric priors. PIC-UIE combines thumbnail-based parameter prediction with a native-resolution UIE executor and uses relative depth only during training.

Learning-based UIE and color representations

  • Learning-based UIE explores perceptual speed, uncertainty, long-range interactions, multiscale detail, and complementary color spaces, but remains dense image-to-image mapping.UIEB is a paired real-image benchmark; examples include FUnIE-GAN (Islam, Xia, and Sattar 2020), uncertainty modeling (Fu et al. 2022), long-range interactions (Peng, Zhu, and Bian 2023), and multiscale detail (Zhang et al. 2024a). UIEC2-Net encodes two color spaces, while Ucolor combines several under transmission guidance (Li et al. 2021); later designs use state-space models, frequency diffusion, dual-domain learning, or task-aware supervision (Guan et al. 2024; Zhao et al. 2024; Peng and Bian 2025; Lin et al. 2026).Source: Related Work · p. 2

Image-adaptive operators and efficient UIE

  • Thumbnail-inferred operators offer resolution-efficient enhancement, while PIC-UIE combines fixed-resolution parameter prediction with a UIE-specific native-resolution executor.Prior approaches use bilateral grids, blended 3D LUTs, spatially varying lookup, and adaptive table sampling (Gharbi et al. 2017; Zeng et al. 2020; Wang et al. 2021a; Yang et al. 2022); these target general photography. Compact UIE networks such as FUnIE-GAN (Islam, Xia, and Sattar 2020) still predict pixels densely, whereas PIC-UIE predicts controls for attenuation, veiling, luminance, coupled chroma, and image-level calibration.Source: Related Work · p. 2

Physical priors and depth supervision

  • Unlike approaches that supply or estimate geometry at inference, PIC-UIE uses relative depth only to regularize its transmission proxy during training.Physical image-formation models describe wavelength- and distance-dependent attenuation and backscatter (Akkaynak and Treibitz 2018, 2019). Related systems use transmission-guided features (Ucolor; Li et al. 2021), RGBD-conditioned diffusion (Osmosis; Bar Nathan et al. 2024), or joint enhancement and depth estimation (Du et al. 2024); the deployed PIC-UIE model is RGB-only, with no depth estimator or branch.Source: Related Work · p. 2

Method

PIC-UIE predicts seven image-adaptive corrections from a fixed 256 × 256 RGB thumbnail and applies them to the native-resolution image through a YCbCr executor. Depth guides transmission-proxy learning during training but is not used by the predictor or executor at inference.

Predictor–Executor Overview

  • At 256 × 256, the predictor produces seven correction outputs grouped into spatial, nonlinear, and calibration components; three stride-2 convolutions create a 32 × 32 feature map for spatial heads.Global average pooling yields a 16-dimensional descriptor, which is combined with a learned projection of thumbnail statistics before the nonspatial heads.Source: Predictor–Executor Overview · p. 2
  • The predictor outputs transmission, veiling-light, and luminance-gain fields, a luminance curve, a coupled chroma LUT, and affine and covariance color-calibration transforms.Three stride-2 convolutions feed spatial heads, while pooled features plus projected thumbnail statistics feed nonspatial heads; native-resolution learned convolution is limited to a 465-parameter luminance refiner.Source: Predictor–Executor Overview · p. 2, Figure 2 · p. 3

Structured YCbCr Corrections

  • In YCbCr, spatial fields correct luminance by subtracting predicted veil and applying a transmission-dependent gain, while a shallow operator refines luminance and a gated 9 × 9 LUT jointly maps chroma.Clipping keeps luminance in range, and the formula avoids division by small transmission values; identity-centered initialization preserves inputs initially.Source: Structured YCbCr Corrections · pp. 2–3
  • The calibration stage applies an affine cross-channel transform followed by blended per-image covariance recoloring, then converts the result from YCbCr back to RGB.The covariance transform uses predicted target mean, lower-triangular recoloring matrix, and blend coefficient; its 3 × 3 eigendecomposition is computed once per image.Source: Structured YCbCr Corrections · p. 3

Depth-Guided Training

  • Depth supervision trains the transmission proxy by correlating −log t with resized depth, enforcing relative ordering while keeping depth out of the predictor and executor.The objective also includes RGB and red-channel reconstruction, SSIM and MS-SSIM, and chroma-LUT smoothness and monotonicity; removing tested reconstruction or LUT terms lowered UIEB PSNR by 0.359–0.456 dB and LSUI PSNR by 0.036–0.173 dB.Source: Depth-Guided Training · p. 3

Experiments · Experimental Setup and Protocol

The experiments evaluate training on UIEB-800, heldout and zero-shot full-reference tests, and separate no-reference and downstream-task datasets. The protocol specifies retrained learning-based baselines, paired and no-reference metrics, training settings, and native-resolution timing conditions.

Experimental Setup and Protocol

  • UIEB-90 is the heldout paired test set after training on UIEB-800, while LSUI is evaluated zero-shot; U45 and UIEB’s unpaired C60 support no-reference testing.UIQAD supplies native-resolution images, while FLSea and URPC2019 support matching and detection, respectively.Source: Experimental Setup and Protocol · pp. 3–4
  • Figure 3 compares qualitative results on UIEB-90 and zero-shot LSUI, with enlarged regions highlighting color correction, local contrast, and structural-detail preservation.Its columns follow the method order in Table 1.Source: Figure 3 · p. 4, Table 1 · p. 4
  • Each learning-based baseline listed in the tables is retrained on UIEB-800 using its official codebase.Table 1 covers full-reference results on UIEB-90 and zero-shot LSUI, with best and second-best cells marked.Source: Experimental Setup and Protocol · p. 4, Table 1 · p. 4
  • Training uses AdamW for 700 epochs with learning rate 3 × 10−4, weight decay 10−4, cosine annealing, batch size 16, and EMA decay 0.999.Paired images are resized to 256×256 and augmented with independent horizontal and vertical flips and random rotations by multiples of 90◦.Source: Experimental Setup and Protocol · p. 4
  • Paired evaluation resizes predictions and references identically to 256 × 256 and reports PSNR, SSIM, MSE, LPIPS-Alex, and PCQI.No-reference evaluation reports UIQM, UCIQE, NIQE, and entropy.Source: Experimental Setup and Protocol · p. 4
  • Native throughput is measured at batch size one in FP32 on an NVIDIA GeForce RTX 4090, using 10 warmups and 30 timed iterations on random tensors, excluding image loading.FLOPs are recomputed with one profiler and counting convention at each table’s listed input resolution; parameter counts use the corresponding model configuration.Source: Experimental Setup and Protocol · p. 4
  • Table 2 assesses zero-shot no-reference quality on U45 and C60, where higher UIQM, UCIQE, and entropy and lower NIQE indicate better scores.Best and second-best values are marked in each column.Source: Experimental Setup and Protocol · p. 3, Table 2 · p. 5

Paired Quality Comparison

PIC-UIE leads UIEB-90 on PSNR and SSIM and has the lowest MSE and LPIPS, while its zero-shot LSUI results are strongest in PSNR and PCQI rather than uniform across all five metrics.

Paired Quality Comparison

  • On UIEB-90, PIC-UIE achieves 24.137 dB PSNR and 0.9216 SSIM, the table’s highest values, and the lowest MSE and LPIPS.Relative to SMDR-IS, it gains 0.427 dB in PSNR and 0.0257 in SSIM, with lower MSE and LPIPS.Source: Paired Quality Comparison · p. 5
  • On zero-shot LSUI without fine-tuning, PIC-UIE records 21.320 dB PSNR with 9.5K parameters and 0.094 GFLOPs; its PSNR and PCQI rank second.Its SSIM, MSE, and LPIPS are outside the top two, so performance is not uniform across all five metrics.Source: Paired Quality Comparison · p. 5
  • Figure 3 shows representative paired comparisons from both test sets, using the same method ordering as Table 1.Source: Paired Quality Comparison · p. 5

No-Reference Quality · High-Resolution Efficiency

PIC-UIE performs competitively on no-reference quality measures while maintaining high native-resolution throughput. Its predictor remains fixed at thumbnail size, but most computation scales with image resolution in the executor.

No-Reference Quality

High-Resolution Efficiency

  • At 1080P and 4K, PIC-UIE runs at 181.9 FPS and 55.0 FPS, requiring 1.83 and 7.20 GFLOPs, respectively.The measurements use the RTX 4090 protocol; dashes represent fixed-size released evaluation pipelines or batch-one FP32 OOM on the 48-GB RTX 4090.Source: High-Resolution Efficiency · p. 5
  • PIC-UIE’s 3.93× increase in FLOPs from 1080P to 4K is close to the fourfold increase in pixel count.Source: High-Resolution Efficiency · p. 5
  • Compared with FUnIE-GAN, the fastest reported native-resolution baseline, PIC-UIE is 2.34× faster at 1080P and 3.14× faster at 4K.Source: High-Resolution Efficiency · p. 5
  • The fixed-size predictor requires 0.038 GFLOPs, or 2.1% of total 1080P computation and 0.5% at 4K, so most computation is in the executor.Source: High-Resolution Efficiency · p. 5
  • PIC-UIE has the highest UCIQE at 1080P and 4K (0.3076 and 0.3203), while FUnIE-GAN has the highest UIQM and introduces visible artifacts in Figure 4 examples.The authors therefore do not treat either metric alone as a perceptual ranking.Source: High-Resolution Efficiency · p. 5

Cross-Hardware Deployment

PIC-UIE was evaluated across six devices using the same FP32 model without architecture changes, with throughput varying substantially by hardware.

Cross-Hardware Deployment

  • The unchanged FP32 model reaches 29.5 FPS at 1080P and 5.9 FPS at 4K on Orin NX, 51.7 FPS at 1080P and 13.3 FPS at 4K on Tesla T4, and 9.1 FPS at 1080P on GTX 750 Ti.Orin NX measurements were averaged over 50 forward passes after warm-up, with inputs resident on the GPU and timing via synchronized CUDA events; the device used MAXN mode with JetPack 6.2, CUDA 12.6, and PyTorch 2.8.0.Source: Cross-Hardware Deployment · p. 5

Downstream Task Evaluation

PIC-UIE ranks first on FLSea local matching and second on URPC2019 detection among the compared enhancement front-ends.

Downstream Task Evaluation

  • 2,185 FLSea matches and 2,078 inliers rank PIC-UIE first, with 95.1% retained after RANSAC; its 0.751 mAP@50 trails NU2Net's 0.766.FLSea matches adjacent frames with SIFT and rejects outliers using RANSAC; URPC2019 trains a separate YOLO11n detector for each enhanced set under the same protocol.Source: Downstream Task Evaluation · p. 5

Structured Design and Capacity

Matched-capacity tests favor structured YCbCr execution over dense predictors, while representation tests favor RGB for prediction and a coupled chroma mapping. Post-hoc capacity sweeps show that widening beyond width 16 trades UIEB gains for LSUI losses and substantially more parameters.

Matched-capacity controls

  • 24.137 dB PSNR for the structured executor exceeds the dense RGB and YCbCr residual predictors’ 20.940 and 20.880 dB, respectively, with 59 additional parameters.The dense predictors differ by only 0.060 dB, versus a 3.197–3.257 dB gap to structured correction; neither the small parameter difference nor a direct YCbCr switch explains the gain.Source: Structured Design and Capacity · pp. 5–6, Table 5 · p. 6

Representation choices

  • RGB is the best predictor input on both datasets, while the coupled 2D LUT outperforms two independent 1D curves and a global affine map.Removing projected RGB statistics causes the largest drop (0.551 dB on UIEB; 0.256 dB on LSUI); direct concatenation adds 1,184 parameters yet remains below the full model. Every tested replacement lowers PSNR on both datasets.Source: Structured Design and Capacity · p. 6, Table 6 · p. 7, Table 7 · p. 7

Capacity scaling

  • 0.201 dB UIEB gain at width 96 comes with 20.2× more parameters and a 0.301 dB LSUI drop; width 160 uses 53.1× more parameters for a 0.117 dB UIEB gain and a 0.338 dB LSUI drop.Widths 8 and 12 trail width 16 on both datasets, and every tested width from 24 to 160 is below width 16 on LSUI. The capacity scan was post hoc; the final architecture and hyperparameters were frozen before evaluation and the sweep was not used for model selection.Source: Structured Design and Capacity · pp. 6–7, Figure 5 · p. 7

Component and Depth Diagnostics

The diagnostics show that executor calibration components contribute most to performance, while depth supervision improves the transmission proxy’s relative ordering without adding inference-time cost. Comparisons across correction representations and predictor widths favor structured outputs over simply increasing model capacity.

Executor components

  • Removing covariance calibration lowers UIEB/LSUI PSNR by 0.910/0.456 dB, and removing the CCM lowers it by 0.730/0.300 dB; these are the largest component-ablation changes.Removing the physical luminance transform, tone curve, chroma LUT, or transmission gate also lowers PSNR on both datasets, but by smaller margins.Source: Component and Depth Diagnostics · p. 7

Depth supervision

  • Depth supervision raises UIEB-90 PSNR from 23.897 to 24.137 dB and LSUI from 21.181 to 21.320 dB, with depth used only in the training loss.Thus, depth adds no inference input, branch, parameters, or deployment-time computation.Source: Component and Depth Diagnostics · p. 7
  • Shuffled depth targets and spatial random noise obtain 23.934 and 23.845 dB on UIEB-90, supporting a benefit from aligned relative depth ordering rather than an arbitrary auxiliary objective.Both alternatives retain the correlation loss but break input alignment.Source: Component and Depth Diagnostics · p. 7

Depth-consistency diagnostic

  • Figure 6 shows mean and median per-image correlations between −log t and depth of 0.729 and 0.785, positive on 88 of 90 images; without depth supervision, the mean is 0.179.The diagnostic supports relative ordering of the transmission proxy, not metric-depth recovery.Source: Figure 6 · p. 7, Component and Depth Diagnostics · p. 7

Overall interpretation

  • Matched dense regressors trail structured correction by 3.197–3.257 dB at similar parameter counts; predictors 20.2×–53.1× wider gain at most 0.201 dB on UIEB and lose at least 0.301 dB on LSUI.The reported controls indicate that changing the correction representation matters more than increasing predictor capacity alone.Source: Component and Depth Diagnostics · p. 7

Conclusion

PIC-UIE predicts structured corrections from a fixed thumbnail and applies them at native resolution in YCbCr, using relative depth only during training. It improves UIEB results over dense RGB and YCbCr residual prediction while running on desktop and edge hardware.

Conclusion

  • At matched capacity, PIC-UIE improves UIEB PSNR by 3.197–3.257 dB over dense RGB and YCbCr residual prediction, reaching 24.137 dB on UIEB-90 and 21.320 dB on zero-shot LSUI.The model has 9,486 parameters.Source: Conclusion · p. 7
  • PIC-UIE estimates corrections from a fixed thumbnail and applies them at native resolution in YCbCr, combining spatial luminance correction, a coupled chroma LUT, and image-level calibration.Relative depth is used only during training.Source: Conclusion · p. 7
  • In FP32, PIC-UIE runs at 55.0 FPS on native 4K inputs on an RTX 4090 and 29.5 FPS at 1080P on an 8-GB Jetson Orin NX; it ranks first for FLSea matching and second for URPC2019 detection.The conclusion reports performance across desktop and edge hardware.Source: Conclusion · p. 7
Loading 2609.33318v1…