Blade Detail Prioritized Exposure Adjustment

• The paper introduces a real-time, geometry-anchored exposure control strategy tailored for UAV-based wind turbine blade inspections.
• It employs LiDAR-derived blade centerline and gimbal-guided projection to lock a blade-centered ROI, ensuring diagnostic brightness across variable conditions.
• Field tests across 120+ flights demonstrate improved brightness statistics and detail preservation, validating the method's effectiveness in both under- and overexposed scenarios.

Searching arXiv for the primary paper and closely related exposure-adjustment work to ground the article in current literature. Searching arXiv for "Automated UAV-based Wind Turbine Blade Inspection: Blade Stop Angle Estimation and Blade Detail Prioritized Exposure Adjustment". Blade Detail Prioritized Exposure Adjustment is a real-time exposure-control strategy introduced for automated UAV inspection of wind turbine blades. In the formulation reported in "Automated UAV-based Wind Turbine Blade Inspection: Blade Stop Angle Estimation and Blade Detail Prioritized Exposure Adjustment," the objective is not generic scene pleasingness but maintaining the blade surface within a diagnostically useful brightness band while preserving fine surface features such as erosion, cracks, and pits during capture rather than after capture (Shi et al., 7 Jul 2025). The method is integrated with a UAV inspection platform, LiDAR-based 3D perception, and gimbal control, and it is motivated by a specific failure mode of conventional auto-exposure: when high-contrast sky and low-albedo blade surfaces occupy varying fractions of the frame, global metering frequently exposes for the background and suppresses blade detail (Shi et al., 7 Jul 2025).

1. Operational context and problem definition

The method is situated in automated wind-turbine inspection, where a UAV traverses along blades of approximately $50\,\mathrm{m}$ length under changing sun angles, cloud cover, and view geometry (Shi et al., 7 Jul 2025). In that setting, conventional auto-exposure is reported to fail because the frame composition is dominated by strongly varying background luminance: bright sky and darker blade surfaces contribute differently as the vehicle moves, so global metering either underexposes the blade or allows overexposure and clipping in blade regions (Shi et al., 7 Jul 2025).

A central premise of the approach is that detail losses incurred at acquisition are irreversible for inspection purposes. The paper states that lost details cannot be restored through post-optimization, and this position defines the method’s emphasis on exposure adjustment during capture rather than post hoc enhancement (Shi et al., 7 Jul 2025). In that sense, Blade Detail Prioritized Exposure Adjustment belongs to the family of acquisition-time metering strategies rather than post-capture exposure correction methods.

The broader exposure-correction literature clarifies this distinction. Multi-exposure fusion methods preserve highlight and shadow content by combining bracketed captures, but they assume multiple aligned exposures and static scenes (Liu, 2022). Single-image correction networks address under- and overexposed sRGB photographs after capture, often through multi-scale decomposition and residual detail reconstruction (Afifi et al., 2020). By contrast, the blade-focused method operates online, inside the sensing-and-control loop of a UAV inspection platform (Shi et al., 7 Jul 2025).

2. Geometry-driven blade locking and ROI construction

The method’s distinguishing mechanism is geometry-driven ROI metering. Rather than selecting a metering window directly in the image, it computes a blade-centered inspection point in 3D from LiDAR, projects that point into the camera, and meters a compact disk around the projection (Shi et al., 7 Jul 2025).

The LiDAR-captured blade point cloud, denoted $\mathcal{P}_b$, is robustly fitted by RANSAC to a line $l_{\mathcal{P}}$, interpreted as the blade centerline (Shi et al., 7 Jul 2025). Let $\mathbf{p}_d$ denote the UAV position. The method constructs the line $l_t$ that passes through $\mathbf{p}_d$ and is perpendicular to $l_{\mathcal{P}}$; the perpendicular foot on the blade centerline is the current 3D inspection point ${}^{w}\mathbf{p}_f$:

$$\mathbf{p}_d \in l_t,\quad l_t \perp l_{\mathcal{P}},\quad l_t \cap l_{\mathcal{P}} = {}^{w}\mathbf{p}_f.$$

This point is then projected into the image using standard pinhole geometry with intrinsics $\mathbf{K}_c$ and extrinsics $\mathcal{P}_b$0:

$\mathcal{P}_b$1

where $\mathcal{P}_b$2 is the depth of $\mathcal{P}_b$3 (Shi et al., 7 Jul 2025).

Given the grayscale image $\mathcal{P}_b$4, the metering region is a disk $\mathcal{P}_b$5 of radius $\mathcal{P}_b$6 centered at $\mathcal{P}_b$7:

$\mathcal{P}_b$8

This construction makes the ROI blade-locked rather than frame-locked. The paper argues that such a design suppresses contamination from sky and ground even when composition changes substantially during traversal (Shi et al., 7 Jul 2025). The gimbal is additionally steered to align with the direction vector of $\mathcal{P}_b$9, keeping the blade center near the image center and reducing foreshortening (Shi et al., 7 Jul 2025). A plausible implication is that the exposure controller and the viewpoint controller are effectively co-designed: the same 3D geometry that stabilizes imaging geometry also stabilizes metering.

3. Camera model, brightness statistic, and deadband control law

The platform uses a DJI M300 UAV equipped with a DJI H20T gimbal camera of $l_{\mathcal{P}}$0, with exposure parameter adjustments performed through the vendor SDK (Shi et al., 7 Jul 2025). The commanded control variable is denoted $l_{\mathcal{P}}$1, described as the current exposure parameter and updated online from the measured blade-ROI brightness (Shi et al., 7 Jul 2025). The paper does not specify which camera degrees of freedom—shutter time $l_{\mathcal{P}}$2, ISO, or aperture $l_{\mathcal{P}}$3—are modified by the SDK in each operating mode.

For completeness, the paper gives the photographic exposure-value relation

$l_{\mathcal{P}}$4

and interprets $l_{\mathcal{P}}$5 conceptually as a proxy for EV, while noting that the mapping from $l_{\mathcal{P}}$6 to hardware settings is handled internally by the camera firmware (Shi et al., 7 Jul 2025).

The feedback statistic is the ROI mean grayscale:

$l_{\mathcal{P}}$7

Given a target brightness band $l_{\mathcal{P}}$8, the controller applies a deadband step update to $l_{\mathcal{P}}$9 (Shi et al., 7 Jul 2025):

• If $\mathbf{p}_d$0, decrease exposure: $\mathbf{p}_d$1.
• If $\mathbf{p}_d$2, increase exposure: $\mathbf{p}_d$3.
• Otherwise, keep $\mathbf{p}_d$4 unchanged.

This law is described as simple and robust, with the deadband preventing oscillation when ROI brightness is already acceptable (Shi et al., 7 Jul 2025). Unlike histogram-based global auto-exposure, the metered luminance is explicitly ROI-weighted by construction:

$\mathbf{p}_d$5

No multi-frame HDR capture, tone mapping, percentile statistics, or highlight-weighted auto-exposure is used in the reported implementation (Shi et al., 7 Jul 2025). That omission is important: the approach solves the blade-inspection problem primarily by metering geometry, not by introducing a more complex photometric estimator.

4. Closed-loop execution and integration with the inspection platform

The full exposure-control loop is executed per frame during inspection (Shi et al., 7 Jul 2025). The reported sequence is:

1. Acquire LiDAR data near the blade and fit the blade centerline $\mathbf{p}_d$6 with RANSAC.
2. Compute the perpendicular line $\mathbf{p}_d$7 from the UAV position to the blade centerline and obtain the foot point $\mathbf{p}_d$8.
3. Drive the gimbal to align with $\mathbf{p}_d$9, then project $l_t$0 into the image to obtain $l_t$1.
4. Form the circular ROI $l_t$2 and compute $l_t$3.
5. Update $l_t$4 by the deadband rule and command the camera exposure (Shi et al., 7 Jul 2025).

The system is embedded in a larger UAV inspection stack that also includes Fermat point based blade stop angle estimation. The paper reports that stop-angle estimation stabilizes hub-center and blade-direction inference, which in turn supports reliable approach geometry and consistent ROI placement on the blade surface (Shi et al., 7 Jul 2025). This suggests that the exposure method is not an isolated camera routine but one component in a geometry-aware autonomy pipeline.

The platform hardware comprises a DJI M300 UAV, a Livox MID-360 LiDAR with $l_t$5 range pitched at $l_t$6, a DJI H20T gimbal camera, and an onboard Intel NUC11TNKi5 with $l_t$7 cores, $l_t$8 threads at $l_t$9, and $\mathbf{p}_d$0 RAM (Shi et al., 7 Jul 2025). The pixel-side computational cost per frame is reported as $\mathbf{p}_d$1, approximately $\mathbf{p}_d$2 for ROI aggregation, while the geometric steps are $\mathbf{p}_d$3 in the number of LiDAR points used by RANSAC (Shi et al., 7 Jul 2025). The method runs in real time on the onboard system alongside navigation (Shi et al., 7 Jul 2025).

A common misunderstanding is to treat the method as a generic computer-vision exposure corrector. In fact, it is inseparable from the platform’s 3D sensing and gimbal actuation: accurate ROI placement depends on blade-centerline recovery, projection, and camera pointing (Shi et al., 7 Jul 2025).

5. Field validation and quantitative effects on blade detail

The reported evaluation spans more than $\mathbf{p}_d$4 flights, $\mathbf{p}_d$5 wind turbine models, and $\mathbf{p}_d$6 operational wind farms, under sunny, cloudy, and overcast conditions, with maximum winds up to $\mathbf{p}_d$7 (Shi et al., 7 Jul 2025). The paper evaluates blade-region grayscale statistics—mean $\mathbf{p}_d$8, standard deviation $\mathbf{p}_d$9, and entropy $l_{\mathcal{P}}$0—as proxies for diagnostically useful brightness and detail richness (Shi et al., 7 Jul 2025).

For the underexposure case, the reported blade-region statistics change as follows (Shi et al., 7 Jul 2025):

Scenario	Before adjustment	After adjustment
$l_{\mathcal{P}}$1	25.48	133.27
$l_{\mathcal{P}}$2	2.48	7.91
$l_{\mathcal{P}}$3	2.10	3.28

The paper further reports that, in the underexposed case, the blade-region histogram mean shifts from $l_{\mathcal{P}}$4 to $l_{\mathcal{P}}$5, and its range expands by $l_{\mathcal{P}}$6 (Shi et al., 7 Jul 2025).

For the overexposure case, the reported blade-region statistics are (Shi et al., 7 Jul 2025):

Scenario	Before adjustment	After adjustment
$l_{\mathcal{P}}$7	245.64	150.85
$l_{\mathcal{P}}$8	5.46	10.25
$l_{\mathcal{P}}$9	1.23	2.96

In the same overexposed scenario, the blade-region histogram mean shifts from ${}^{w}\mathbf{p}_f$0 to ${}^{w}\mathbf{p}_f$1, and its range expands by ${}^{w}\mathbf{p}_f$2 (Shi et al., 7 Jul 2025).

Qualitatively, the paper states that details invisible in underexposed frames become legible after adjustment, while blown highlights on overexposed upper blade surfaces are driven back toward a midtone regime where fine texture is discernible (Shi et al., 7 Jul 2025). The significance of these numbers is not merely photometric normalization. In the paper’s interpretation, the increases in ${}^{w}\mathbf{p}_f$3 and ${}^{w}\mathbf{p}_f$4 indicate expanded tonal variation and richer blade detail within the region relevant for inspection (Shi et al., 7 Jul 2025).

6. Position within exposure-adjustment research

Blade Detail Prioritized Exposure Adjustment occupies a specific point in the exposure-adjustment design space. It is a real-time, acquisition-stage, geometry-anchored ROI metering method (Shi et al., 7 Jul 2025). Related work spans several different regimes.

Multi-exposure fusion preserves shadow and highlight content by combining multiple bracketed frames; the "Perceptual Multi-Exposure Fusion" method improves adaptive well-exposedness in YCbCr and uses a 3-D gradient to extract fine details, while explicitly targeting lower complexity than detail-enhanced fusion methods (Liu, 2022). Such methods are relevant when static scenes and bracketed captures are available, but the blade-inspection system described in the UAV paper does not use multi-frame HDR or fusion (Shi et al., 7 Jul 2025).

Learning-based single-image exposure correction addresses post-capture over- and underexposure. "Learning Multi-Scale Photo Exposure Correction" formulates correction as separate low-frequency color enhancement and high-frequency detail enhancement using a Laplacian pyramid and a coarse-to-fine DNN (Afifi et al., 2020). "Curve Distillation for Efficient and Controllable Exposure Adjustment" instead uses a lightweight student network to predict a spatially varying affine map ${}^{w}\mathbf{p}_f$5 conditioned on an exposure map, enabling local controllability without paired training data (Li et al., 2022). "Practical Exposure Correction: Great Truths Are Always Simple" proposes a linear compensation rule with an exposure-sensitive adversarial function and a segmented shrinkage iterative scheme, emphasizing efficiency and robustness (Ma et al., 2022). These methods operate on already captured images, whereas the blade-focused UAV method attempts to prevent irreversible loss of detail at acquisition time (Shi et al., 7 Jul 2025).

Adaptive metering and control have also been studied directly. "Personalized Exposure Control Using Adaptive Metering and Reinforcement Learning" models exposure control in live preview as an MDP with a fully convolutional network that predicts continuous ${}^{w}\mathbf{p}_f$6 from the current frame and an adaptive metering head (Yang et al., 2018). This is the closest adjacent line of research conceptually, because both methods modify exposure online rather than retrospectively. The UAV method differs in that its metering map is not learned from semantics; it is derived from 3D blade geometry and projected into image space (Shi et al., 7 Jul 2025).

More recent semantic-aware correction methods move further toward object-conditioned optimization. "CLIP-Guided Unsupervised Semantic-Aware Exposure Correction" injects FastSAM-derived semantics into a multi-scale residual spatial Mamba network and uses CLIP-guided pseudo-ground truth for unsupervised training (Wu et al., 27 Jan 2026). "WEC-DG: Multi-Exposure Wavelet Correction Method Guided by Degradation Description" uses degradation descriptors, wavelet-domain light-detail decoupling, and dual Exposure Consistency Alignment Modules for exposure restoration and detail reconstruction (Zhao et al., 13 Aug 2025). This suggests a broader taxonomy: geometry-aware online metering at capture time, semantics-aware online control, and semantics-aware or wavelet-aware post-capture correction occupy complementary rather than interchangeable roles.

7. Limitations, failure modes, and prospective extensions

The reported method uses only the mean ROI brightness ${}^{w}\mathbf{p}_f$7 as its feedback variable (Shi et al., 7 Jul 2025). It does not explicitly optimize edge energy, gradient magnitude, Laplacian structure, entropy, or highlight clipping during control. The paper therefore identifies a limitation: the controller regulates brightness, not detail directly (Shi et al., 7 Jul 2025). A plausible implication is that improved detail preservation arises indirectly through better blade-tone placement rather than direct optimization of texture or crack visibility.

Several failure modes are explicitly noted. ROI contamination can occur if the projected foot lies too close to blade edges, allowing sky pixels into ${}^{w}\mathbf{p}_f$8 (Shi et al., 7 Jul 2025). Extreme backlighting or specular glare may saturate the ROI, in which case the deadband controller can reduce ${}^{w}\mathbf{p}_f$9 but cannot recover detail once sensor dynamic range is exceeded (Shi et al., 7 Jul 2025). In very low light, ISO and shutter limits may induce noise or motion blur, which simple brightness control cannot correct (Shi et al., 7 Jul 2025). The method also assumes sufficient stabilization for reliable projection and ROI tracking, although the platform operates at speeds no greater than $$\mathbf{p}_d \in l_t,\quad l_t \perp l_{\mathcal{P}},\quad l_t \cap l_{\mathcal{P}} = {}^{w}\mathbf{p}_f.$$0 near the blade to reduce motion blur (Shi et al., 7 Jul 2025).

The paper outlines several future directions. It notes that learning-based metering or multi-objective optimization could push detail preservation further, that HDR bracketing or burst fusion could address extreme dynamic range, that explicit joint control of shutter, ISO, and aperture would manage blur-noise trade-offs more directly, and that PID, adaptive, or predictive controllers using both brightness error and detail cues could improve convergence and stability (Shi et al., 7 Jul 2025). Adjacent literature supplies concrete exemplars for such extensions: adaptive metering learned by reinforcement learning (Yang et al., 2018), multi-exposure fusion for highlight and shadow detail (Liu, 2022), and descriptor- or semantic-guided detail-aware correction in wavelet or vision-language frameworks (Zhao et al., 13 Aug 2025, Wu et al., 27 Jan 2026).

In summary, Blade Detail Prioritized Exposure Adjustment denotes a geometry-anchored, ROI-weighted, real-time exposure controller for UAV blade inspection, not a generic image-enhancement operator. Its central technical contribution is to replace scene-global metering with blade-locked metering derived from LiDAR geometry and camera projection, and its practical importance lies in shifting exposure correction upstream to the moment of capture, where inspection-critical detail can still be preserved (Shi et al., 7 Jul 2025).