U3LIE: UAV Low-Light Enhancement for UHR Imagery

• U3LIE is a lightweight, unsupervised framework tailored for ultra-high-resolution UAV low-light image enhancement with real-time 4K performance.
• It integrates a training-only Adaptive Pre-enhancement Augmentation with a curve-based network and patch-wise Luminance Interval Loss to balance exposure.
• Empirical results show U3LIE outperforms competitors in Edge Efficiency Index and speed, emphasizing its practical deployability in challenging UAV scenarios.

Searching arXiv for U3LIE and closely related low-light enhancement work to ground the article. U3LIE is an end-to-end, zero-reference enhancement framework tailored for Ultra-High-Resolution (UHR) UAV low-light images. It was introduced together with the U3D benchmark and the Edge Efficiency Index (EEI) to address a specific deployment regime in which existing low-light image enhancement methods struggle with Ultra-High Resolution, lack of paired data, severe non-uniform illumination, and on-board resource constraints. Its central design couples a lightweight curve-based enhancement network with two training-only mechanisms—Adaptive Pre-enhancement Augmentation (APA) and the Luminance Interval Loss $\mathcal{L}_{int}$—while preserving real-time full-4K inference without patching (Lu et al., 1 Sep 2025).

1. Problem setting and system objective

U3LIE targets unsupervised UHR UAV low-light image enhancement rather than paired image-to-image restoration. The motivating setting is aerial night imagery, where sensor noise, color casts, and strong local illumination disparities interact with strict throughput and memory limits. Within that regime, the framework is explicitly optimized not only for perceptual enhancement but also for deployment factors including speed, resolution, model complexity, and memory footprint, which are formalized in the accompanying EEI metric (Lu et al., 1 Sep 2025).

The framework is described as a holistic component of a broader system comprising dataset, metric, and method. In that system, U3LIE is the method layer: a lightweight enhancement model intended to operate directly on raw low-light UAV images at native 4K resolution during inference. A plausible implication is that the method is organized around deployment realism rather than laboratory-only image quality optimization, since the paper treats real-time throughput and VRAM usage as first-order criteria alongside no-reference perceptual metrics.

2. Pipeline and architectural design

The U3LIE pipeline is divided into training and inference phases. During training, APA transforms each raw low-light frame $I_{in}$ into a moderately corrected proxy $I_{aug}$; a lightweight enhancement network $\mathcal{F}_{net}$ then processes $I_{aug}$ to predict a pixel-wise adjustment map $A$; and an enhancement operator $\mathcal{C}^{(N)}$, described as an iterative, higher-order curve adjustment, combines $A$ with the original $I_{in}$ to produce the enhanced image $I_{enh}$. Training minimizes a compound objective formed by $I_{in}$0 together with smoothness, spatial coherence, and color-constancy terms. During inference, APA is disabled, and the network processes raw $I_{in}$1 directly to produce $I_{in}$2 at full 4K resolution in real time (Lu et al., 1 Sep 2025).

The backbone is derived from the curve-based DCE++ design. For UHR efficiency, all down-sampling and up-sampling layers are removed so that the model operates at native resolution, and the internal feature width is reduced from 32 channels to 8 channels. The network is described as a small stack of $I_{in}$3 convolutions with ReLU activations, ending in an 8-channel adjustment map $I_{in}$4. On 4K input, the reported compute and memory footprint are 1.32 K parameters, 9.91 GFLOPs, and 3.23 GB VRAM (Lu et al., 1 Sep 2025).

Component	Role	Reported detail
APA	Training-only pre-processor	Produces $I_{in}$5 from $I_{in}$6
$I_{in}$7	Lightweight enhancement network	Predicts 8-channel map $I_{in}$8
$I_{in}$9	Enhancement operator	Iterative, higher-order curve adjustment

A recurring point of clarification is that APA is not part of the deployed inference path. The paper states that the APA module is disabled at inference, so U3LIE’s real-time behavior depends on the lightweight network and curve-based operator rather than on the preprocessing stack.

3. Adaptive Pre-enhancement Augmentation

APA is a training-only, one-time pre-processor $I_{aug}$0 intended to stabilize the extreme dynamic range and color casts of UAV night images. It consists of three stages. The first stage performs edge-preserving denoising and color-space separation: a bilateral filter $I_{aug}$1 is applied to $I_{aug}$2, after which the denoised BGR image is converted into YCrCb space to isolate luminance $I_{aug}$3 and chroma $I_{aug}$4 (Lu et al., 1 Sep 2025).

The second stage performs adaptive luminance and contrast boosting. APA computes the normalized mean luminance as

$I_{aug}$5

then derives an adaptive gamma

$I_{aug}$6

Gamma correction and CLAHE are then applied: $I_{aug}$7 after which a BGR image $I_{aug}$8 is reconstructed from $I_{aug}$9 (Lu et al., 1 Sep 2025).

The third stage performs perceptual color and highlight correction. In Lab space, the red channel is adjusted according to

$\mathcal{F}_{net}$0

and the result is converted back to BGR, yielding $\mathcal{F}_{net}$1. In HSV space, saturation and value are rescaled as

$\mathcal{F}_{net}$2

producing the final augmented image $\mathcal{F}_{net}$3 (Lu et al., 1 Sep 2025).

The stated function of APA is normalization rather than final enhancement. This suggests that its principal contribution is to regularize the optimization landscape seen by $\mathcal{F}_{net}$4, especially under the severe illumination variability of UAV night data, while preserving the simplicity of the deployed inference graph.

4. Luminance Interval Loss and optimization objective

The Luminance Interval Loss $\mathcal{F}_{net}$5 is introduced to jointly avoid under-exposure and over-exposure. The enhanced image $\mathcal{F}_{net}$6 is partitioned into non-overlapping patches $\mathcal{F}_{net}$7. If $\mathcal{F}_{net}$8 denotes a patch mean luminance and $\mathcal{F}_{net}$9 the global mean,

$I_{aug}$0

then the local dark, local bright, and global terms are defined as

$I_{aug}$1

$I_{aug}$2

$I_{aug}$3

The combined loss is

$I_{aug}$4

The paper gives $I_{aug}$5 as a lower luminance threshold, $I_{aug}$6 as an upper threshold, $I_{aug}$7 as a target global mean, and $I_{aug}$8 as a balance parameter; the example values are $I_{aug}$9, $A$0, $A$1, and $A$2 (Lu et al., 1 Sep 2025).

The total training objective is

$A$3

where $A$4 is a spatial-coherence loss, $A$5 is TV smoothness on $A$6, and $A$7 enforces color constancy. The reported weights are $A$8, $A$9, $\mathcal{C}^{(N)}$0, and $\mathcal{C}^{(N)}$1; the $\mathcal{C}^{(N)}$2 patch size is $\mathcal{C}^{(N)}$3 (Lu et al., 1 Sep 2025).

The design intent is explicit: $\mathcal{C}^{(N)}$4 enforces each patch’s luminance to lie in $\mathcal{C}^{(N)}$5 while nudging the overall mean toward $\mathcal{C}^{(N)}$6. A plausible implication is that the loss directly constrains exposure statistics at both local and image-wide scales, which is particularly relevant in aerial scenes containing simultaneously dark ground regions and saturated artificial light sources.

5. Training protocol and empirical results

Training is performed on $\mathcal{C}^{(N)}$7 image patches with batch-size 4 and 4-step gradient accumulation, yielding an effective batch size of 16. The optimizer is Adam with weight decay $\mathcal{C}^{(N)}$8. The learning rate uses linear warmup for 5 epochs to $\mathcal{C}^{(N)}$9, is halved at epochs 50 and 100, and training lasts 100 epochs. Gradient clipping uses a maximum norm of 0.05. The implementation is in PyTorch on Ubuntu 20.04 with a single NVIDIA RTX 3090 (24 GB) and random seed 2025 (Lu et al., 1 Sep 2025).

For 4K inference at $A$0, the reported speed is 23.80 FPS on RTX 3090 without patching. The model cost is reported as 1.32 K parameters, 9.91 G FLOPs, and 3.23 GB VRAM. On no-reference perceptual metrics, U3LIE reports NIQE $A$1, BRISQUE $A$2, and PI $A$3. Its EEI is $A$4, with lower values indicated as better (Lu et al., 1 Sep 2025).

The paper’s comparison against state-of-the-art methods emphasizes the interaction between efficiency and quality. U3LIE is reported to achieve the best EEI by a large margin: DCE++ has EEI $A$5, RUAS $A$6, and ZERO-DCE $A$7. In speed, the reported comparison is 23.8 FPS for U3LIE versus 7.76 FPS for ZERO-DCE, 7.64 FPS for RUAS, and 7.68 FPS for DCE++. Heavy GAN and Transformer methods are described as either out-of-memory on 4K or running at less than 1 FPS. In perceptual quality, U3LIE’s PI $A$8 is reported to outperform other real-time models and to rival larger unpaired GANs, with CycleGAN given as PI $A$9 but EEI $I_{in}$0 (Lu et al., 1 Sep 2025).

Ablation results isolate the contributions of APA and $I_{in}$1. The baseline DCE++ reports PI $I_{in}$2; adding APA alone yields PI $I_{in}$3; adding $I_{in}$4 alone yields PI $I_{in}$5; and combining APA with $I_{in}$6 yields PI $I_{in}$7. Hyperparameter sweeps are reported to confirm that the interval $I_{in}$8 and $I_{in}$9 strike the best perceptual-fidelity balance (Lu et al., 1 Sep 2025).

6. Interpretation, deployment significance, and common points of confusion

U3LIE is best understood as a deployment-oriented zero-reference LIE framework rather than as a generic image enhancement network. Its defining features are not only unsupervised training and curve-based adjustment, but also the separation between training stabilization and inference efficiency: APA is used only during optimization, while the runtime path remains lightweight. This distinguishes the training procedure from the deployed system and counters the common misunderstanding that the reported 4K real-time performance includes the full APA preprocessing stack (Lu et al., 1 Sep 2025).

The method’s evaluation framework also shapes its interpretation. Because EEI balances perceptual quality with speed, resolution, model complexity, and memory footprint, U3LIE is not optimized solely for a single image-quality criterion. This suggests that direct comparison to methods that are out-of-memory on 4K or operate below real-time throughput must be understood within the paper’s stated deployment constraints rather than as a purely aesthetic ranking.

Qualitatively, the reported behavior is specific: U3LIE is said to recover detail in shadows without clipping bright streetlights or introducing color casts. Distribution plots reportedly show that it centers the luminance distribution in a perceptually pleasing mid-tone, maximizes contrast and entropy, and keeps sharpness moderate so as to avoid noise amplification (Lu et al., 1 Sep 2025). Within the paper’s scope, these observations reinforce the stated role of APA and $I_{enh}$0: the former regularizes difficult raw inputs during training, and the latter constrains exposure statistics so that enhancement remains balanced across locally heterogeneous aerial night scenes.

In summary, U3LIE denotes a lightweight, end-to-end, zero-reference framework for unsupervised UHR UAV low-light enhancement that combines a DCE++-derived native-resolution backbone, training-only Adaptive Pre-enhancement Augmentation, and a patch-wise/global Luminance Interval Loss. Its reported significance lies in the joint optimization of visual quality, full-4K throughput, and deployability under UAV-specific constraints (Lu et al., 1 Sep 2025).