Acquisition Inspection

• Acquisition inspection is a discipline that ensures data-acquisition processes yield reliable measurements through specification, control, and verification.
• It encompasses the design of specialized sensor architectures and active calibration, synchronization, and quality assurance for defect detection.
• The approach integrates closed-loop sensing and rigorous configuration management to maintain optimal performance across diverse inspection domains.

Taken together, the cited work suggests that acquisition inspection denotes the specification, control, verification, and monitoring of data-acquisition processes whose outputs must be reliable enough for downstream inspection, diagnosis, or measurement. In this literature, the concept spans both acquisition for inspection—for example, close-range multimodal capture of rotor blades, sewer interiors, photovoltaic arrays, or road surfaces—and inspection of acquisition itself, including camera alignment, synchronization, DAQ configuration management, acquisition accuracy evaluation, and acquisition-state drift monitoring in AI systems (Wittmann et al., 19 Jun 2026, 0803.0194, Ying et al., 2020, Soliman, 11 Jun 2026).

1. Scope and conceptual structure

The corpus uses the concept in at least two closely related senses. In the first, acquisition inspection is the engineering of sensing pipelines so that the captured data support later defect detection, localization, classification, and quantification. Examples include a UAV-mounted multimodal platform for wind-turbine blades operating at close range with dense 3D geometry, RGB, and thermal IR in one common metric frame, a structured-light sewer inspection system that produces cylindrical texture and displacement maps, and a visual-servoed UAV that tracks the middle of photovoltaic arrays to obtain detailed low-altitude imagery (Wittmann et al., 19 Jun 2026, Künzel et al., 2023, Velasco-Sánchez et al., 2023).

In the second sense, acquisition inspection concerns the inspection of the acquisition chain itself. Here the object of inspection is not the asset under test but the sensing or DAQ system: acquisition accuracy of visual boards is characterized by internal noise, ADC quantisation behaviour, analogue processing, dominant frequencies, and synchronisation accuracy; astronomical acquisition cameras are aligned and verified against encircled-energy and flexure requirements; distributed fusion DAQ infrastructures are supervised through centralized consoles; and CT-based AI systems are checked against an acquisition envelope that is measurable from pixels but not necessarily from DICOM metadata (0803.0194, Araiza-Duran et al., 2024, Ying et al., 2020, Soliman, 11 Jun 2026).

This suggests that acquisition inspection is best understood as a layered discipline. One layer concerns what must be sensed for inspection to be meaningful; another concerns how the sensing system is geometrically, temporally, and operationally constrained; and a third concerns how the resulting acquisition state is validated and monitored before downstream inference or human interpretation.

2. Inspection objectives and information requirements

Across domains, acquisition design is driven by the defect physics and by the decision variables that downstream inspection must recover. For wind-turbine rotor blades, the stated objective is predictive maintenance through high-quality close-range data that support detection and localisation of surface anomalies, reliable classification of defect types, and quantification of defect geometry—size, depth, and shape—at millimetre scale; the same system explicitly targets dense 3D geometry, high-resolution RGB, and thermal IR so that anomalies visible in one modality can be localized and quantified in 3D and cross-checked in the others (Wittmann et al., 19 Jun 2026).

For sewer inspection, the informational target is similarly geometric. The structured-light system is designed to detect and classify spatial defects such as jutting intrusions, spallings, misaligned joints, bent pipes, and local shape defects, and it converts fused 3D data into a cylindrical base mesh, a texture map, and a displacement map. In that representation, radial deviation is expressed as $\Delta r(\phi,z)=\rho(\phi,z)-r_0$, which directly supports manual measurement and future automatic defect detection (Künzel et al., 2023).

Other domains formalize acquisition objectives differently but with the same logic. Grain appearance inspection is formulated as anomaly detection because healthy kernels are abundant while damaged grains and impurities are diverse and open-set; the acquisition device is therefore built to expose as much kernel surface as possible in a consistent imaging geometry (Fan et al., 2023). In remote sensing, image-quality inspection is explicitly positioned between acquisition and application: clouds, shadows, blur, stitching defects, stripe noise, and pixel missing must be localized and their area share computed so that data can be repaired or reacquired if necessary (Yu et al., 2023). In lung-nodule AI, the problem is framed as whether incoming studies remain within the acquisition envelope on which the detector was validated; here the relevant variables are the frequency/kernel axis and the noise axis, which map to different failure modes—measurement instability versus detection fragility (Soliman, 11 Jun 2026).

The recurring principle is that acquisition is not neutral. The sensed variables, standoff distance, spatial resolution, spectral channels, and physical viewpoint determine which defects are even observable, and which are measurable rather than merely visible.

3. Sensor architectures and multimodal acquisition systems

A central pattern in the literature is the construction of application-specific sensor architectures rather than generic camera rigs. The wind-blade platform combines an industrial RGB camera, a passive thermal infrared camera, and an in-house 3D scanner based on two $2/3"$ global-shutter monochrome stereo cameras with a $\approx 1$ m baseline and a $2$ W, $520$ nm laser speckle projector. All sensors are rigidly mounted, co-calibrated to one stereo camera, and synchronized so that every 3D point can be assigned RGB colour and thermal intensity in a common metric frame (Wittmann et al., 19 Jun 2026).

The sewer system adopts a different but equally deliberate geometry: six identical single-shot structured-light modules are arranged around a cylindrical carrier so that their fields of view overlap and provide 360° coverage of pipes with 200–400 mm diameter. Each module contains a camera and projector with parallel optical axes and a 29° stereo angle, and the acquisition alternates texture–3D–texture images to support both geometry and appearance reconstruction under motion (Künzel et al., 2023).

In agricultural inspection, the AI4GrainInsp prototype uses two vertically aligned industrial cameras, one above and one below a transparent plate, to capture both sides of cereal kernels. With conveyor-based presentation and vibration-assisted separation, the device attains 92–98% superficial area coverage of each kernel, which is a direct acquisition answer to the fact that small defects may appear on any side (Fan et al., 2023).

Acquisition architectures need not be optical alone. In the CLIC module work, the acquisition system is organized as one local crate per module with one service board and several standard motherboards, each carrying mezzanines for subsystem-specific front ends. In that setting, acquisition inspection includes radiation tolerance, optical GBT links, jitter control, modularity, and slow-control monitoring via GBT SCA rather than image formation alone (1111.7176).

These examples show that modality choice, baseline, field of view, mounting rigidity, and physical packaging are not secondary implementation details. They define the feasible inspection observables.

4. Calibration, synchronization, and acquisition quality assurance

Once the sensor architecture is fixed, acquisition inspection becomes a problem of geometric, temporal, and radiometric trustworthiness. The wind-blade system calibrates intrinsics and extrinsics with Zhang’s method in OpenCV, using a glass checkerboard for RGB and monochrome cameras and a heated FR4 PCB checkerboard for the thermal camera. All sensors are referenced to one stereo camera, and the paper reports individual reprojection errors below $0.2$ px, with the standard projection model

$\tilde{x}_i = K_i(R_iX+t_i)$

used implicitly for multimodal projection and colorisation (Wittmann et al., 19 Jun 2026).

Temporal integrity is treated just as explicitly. The same platform uses an Arduino-based trigger board that hardware-triggers stereo cameras, RGB camera, laser driver, and piezo driver; generates image timestamps with $4\,\mu$s resolution corresponding to the centre of exposure; and pairs the continuously streaming thermal camera in software by nearest timestamp. A typical cycle acquires 8–16 stereo pairs plus one RGB frame, with 10 stereo pairs plus one RGB frame taking 117 ms in laboratory tests (Wittmann et al., 19 Jun 2026).

Where the object of inspection is the acquisition device itself, the metrics are broader. The visual-inspection-board evaluation paper organizes acquisition accuracy into five parameter groups: internal noise, video ADC quantisation parameters, analogue processing section parameters, dominant frequencies, and synchronisation accuracy. Its synchronisation indicator,

$$Sy=\frac{1}{N}\sum_{k=0}^{N-1}\sum_{q=1}^{Q}|m_q(k)-M_q|,$$

is an image-based measure of lock-in jitter in pixel units, complementing noise and black-level metrics that are directly relevant to metrologic inspection (0803.0194).

Astronomical acquisition imposes a different validation regime. For the SOXS acquisition and guiding camera, the central requirement is at least 80% geometric encircled energy within two pixels over the central field. The alignment campaign uses tolerance analysis, Monte Carlo simulation, focus optimization, distortion tests, and edge-based FWHM estimation; 92% of Monte Carlo cases satisfy the 80% EE criterion, and measured FWHM remains below two pixels across the usable field (Araiza-Duran et al., 2024).

A further extension is acquisition-state validation for AI. In lung-nodule CT, a four-feature pixel fingerprint recovers reconstruction identity with patient-level AUC about 0.95 on real CT and 0.995 on a QIBA phantom, even when the DICOM ConvolutionKernel tag is uninformative. The practical implication is that acquisition quality assurance cannot always be delegated to metadata; input-side validation may need to operate directly on pixels (Soliman, 11 Jun 2026).

5. Planning, control, and active acquisition

A large part of contemporary acquisition inspection is the active selection of where, when, and how to sense. In robot manipulator inspection, point clouds from an eye-in-hand RealSense D435 are converted into object profiles, normals are estimated, and target poses are generated so that the camera principal axis aligns with the local surface normal. Target spacing and stand-off distance are then filtered before the poses are passed to MoveIt/OMPL for execution; the method supports both single-path and multi-path inspection plans for objects of varying shape and scale (Tasneem et al., 2023).

For PV arrays, visual servoing and NMPC are used to replace photogrammetric grid flight with image-driven acquisition. Two image points defining the array midline are converted into line features $(^{i}r,{}^{i}\theta)$, a visual-servoing law computes a body-velocity command, and a constrained NMPC tracks that command while enforcing altitude and velocity limits. The explicit purpose is to keep the middle of the underlying PV array centered during low-altitude flight so that the UAV acquires detailed images rather than high-altitude orthomosaic input with large amounts of useless data (Velasco-Sánchez et al., 2023).

In cluttered GNSS-denied environments, the LiDAR-based quadrotor system adopts a dual-phase workflow. During human-in-the-loop inspection, untrained pilots specify semantically meaningful viewpoints while IPC-based local planning, Safe Flight Corridors, and MPC enforce obstacle avoidance. During autonomous inspection, the recorded inspection points are reordered by a Traveling Salesman Problem using A* path lengths on a global occupancy grid as edge costs, then executed automatically. Field experiments report trajectory-length reductions of up to 40% and flight-time reductions of 57% relative to the human phase (Liu et al., 29 Mar 2025).

Active acquisition also appears outside robotics. In single-beam SEM, a fast low-fidelity scan is followed by learning-guided selection of a sparse subset of pixels for high-resolution rescanning, using an error estimator and a weighted determinantal point process to balance saliency and spatial diversity; the paper reports speedups of up to an order of magnitude (Mi et al., 2021). In underwater inspection, active learning uses mutual information with Monte Carlo dropout to select informative frames for segmentation, reaching 67.5% meanIoU with 12.5% of the data on a pipeline dataset, compared with 61.4% for random selection at the same label budget (Marnet et al., 2024). In spacecraft inspection, reinforcement learning uses Sun-aware rewards so that only well-illuminated points count as inspected, yielding an interquartile mean percentage of inspected points of 98.82% for the finalized model (Wijk et al., 2023).

These works indicate that acquisition inspection increasingly includes closed-loop sensing policies rather than static capture scripts.

6. Configuration management, monitoring, and operational governance

Inspection-grade acquisition depends not only on sensors and controllers but also on configuration discipline. In the EAST fusion experiment, the data acquisition console is the central supervisory component for around 60 DAQ nodes and more than 3000 channels. The upgraded LabVIEW-based console organizes configuration into Status, DAS, Card, Control, and Monitor views; supports add/delete node operations, Excel-based download/upload, configuration backup, parameter publishing, log query, and shot simulation data acquisition test; and exposes node states via a color-coded monitor view (Ying et al., 2020).

The CLIC module work addresses a related problem under radiation and tunnel constraints. Its local crate architecture, point-to-point GBT optical links, modular motherboards and mezzanines, and generic acquisition mezzanine are explicitly designed for inspectability, maintainability, and qualification. Monitoring is further extended by the planned use of GBT SCA for ADCs, DACs, JTAG, I²C, SPI, and alarm monitoring, enabling online inspection of crate voltages, temperatures, and link status (1111.7176).

In remote sensing image quality inspection, operational governance is realized as a two-step deep-learning system: SwinV2 first performs multi-label block classification, then region-like defects are localized by SegFormer while pixel-like defects are handled by threshold-based image processing. The system exists specifically because quality inspection is “an indispensable step between the acquisition and the application” of remote sensing images, and because pass/fail decisions depend on explicit localization and area share of defects rather than on a scalar quality score (Yu et al., 2023).

Medical-imaging governance pushes this logic further. The CT acquisition-state study argues that local acceptance testing and ongoing drift monitoring need an input-side layer that checks whether incoming studies remain inside the acquisition envelope on which a model was validated. Because kernel-driven measurement instability and noise-driven detection fragility are partly invisible to metadata, acquisition-aware monitoring becomes a prerequisite for trustworthy AI deployment (Soliman, 11 Jun 2026).

7. Limitations and emerging directions

The literature is explicit that many current systems remain transitional. The wind-blade multimodal platform is still at laboratory validation stage; airborne trials, trajectory-based motion compensation, shorter exposure times, and automated defect detection and classification remain future work, and direct sunlight is noted as a limiting factor for the optical system (Wittmann et al., 19 Jun 2026). The geometry-driven robot path planner for visual inspection has no explicit global coverage guarantee, depends on depth quality, and still requires manual cluster selection (Tasneem et al., 2023). The road-digital-twin framework improves perception and decision-making, but the simulator still exhibits a sim-to-real gap, and defect libraries remain finite (Zhang et al., 2024).

Several papers also indicate that acquisition inspection is moving toward richer closed loops. In PV inspection, thermal/RGB fusion is coupled to adaptive re-acquisition via Rodrigues-based gimbal updates, while geospatial de-duplication and relevance-only telemetry reduce duplicate alerts and transmitted data (Lysyi et al., 6 Dec 2025). In CT AI, acquisition-aware, input-side validation is presented as the missing layer beneath output monitoring and metadata auditing (Soliman, 11 Jun 2026). This suggests a broad future direction in which acquisition is no longer treated as a static precondition for inspection, but as a measurable, controllable, and continuously audited variable within the inspection system itself.