Image-based fault monitoring in additive manufacturing, driven by per-layer cameras paired with deep learning, can detect warping, delamination, extrusion failure, and powder-bed anomalies within the layer where the defect originates, a 2023 survey of 53 papers concludes [S1].
The work covers metal powder-bed fusion (PBF) and polymer fused-deposition modeling (FDM) systems alike, and notes that in-line vision is now positioned as the primary quality gate before post-process CT or destructive metallography [S1]. Process engineers should treat layer imaging as a real-time process-control sensor, not a post-build inspection tool.
What Layer Imaging Actually Measures
A calibrated camera at a known pose inside the build chamber captures the as-printed surface after each new layer is deposited, and the image is segmented into a binary mask of the current part silhouette [S4]. The system then renders a synthetic image of what the CAD model predicts for that exact layer height, applies perspective projection, and computes a pixel-wise deviation map between predicted and observed geometry [S4].
Three failure modes are statistically detectable from a single camera and one light source: workpiece warpage (edge curl away from the build plate), build plate delamination (full or partial lift-off of the part), and extrusion failure (missing or skipped tracks) [S4]. The 2017 DTU / euspen study demonstrated unambiguous detection of the failure time point for all three modes on a low-cost FDM printer, using only well-known image-processing primitives [S4].
A 2025 review describes a high-resolution static analysis workflow where every layer is stored as a full image, enabling offline re-inspection of any layer in the build after a defect flag fires, rather than discarding image data in real time [S2].
Comparison of Layer-Image Approaches
The main implementation paths differ in four decision criteria: trigger source, image modality, model dependency, and best-fit process [S1][S3][S5][S8].
1) Trigger source. Per-layer triggered capture (one image after each recoat or layer change) versus continuous video. Per-layer capture keeps data volume bounded by layer count; continuous video requires frame differencing between consecutive layers to isolate the new geometry [S1][S4].
2) Image modality. Visible-light RGB cameras dominate FDM work because extrusion defects and warpage are geometric; thermal-infrared cameras are layered on PBF to capture melt-pool thermal signatures that precede geometric defects; electron-beam imaging applies to EBM systems where optical access is blocked by the electron column [S5][S8].
3) Model dependency. CAD-to-image comparison (model-based) requires accurate build-plate and camera calibration but generalises across parts without training data; deep-learning classifiers (data-driven) need labelled defect image sets per process but tolerate geometric ambiguity and process drift better [S1][S3][S4].
4) Best-fit process. FDM and material extrusion benefit most from single-camera geometric comparison because the part silhouette changes predictably each layer; PBF benefits from thermal imaging plus visible-light overlay because subsurface defects can be inferred from surface thermal transients before they become geometric [S4][S8]. A 2026 FFF study combined RGB imaging with multimedia sensor data to predict extrusion failure from time-series features rather than single-layer snapshots [S6].
Where Layer Imaging Pays Off, and Where It Falls Short
For FDM and polymer material extrusion, layer imaging pays off because the dominant failure modes are surface-visible: warpage, delamination, and missing extrusion are all observable in a single overhead image within one layer of occurrence [S4].
For PBF metal systems, layer imaging pays off when combined with thermal or melt-pool imaging, since powder-spread defects and lack-of-fusion porosity can appear as surface texture anomalies or as delayed cool-down signatures in the infrared band [S8].
The 2023 PMC survey notes that published image-based monitoring is still dominated by metal-PBF and FDM case studies; binder jetting, directed energy deposition with multi-axis heads, and large-format polymer extrusion remain under-covered in the public literature [S1].
Layer imaging is not a substitute for volumetric inspection: subsurface porosity, residual stress, and grain-structure defects are not visible from a single top-down image and still require CT scanning or metallographic sectioning for confirmation [S1].
Hardware Stack and Sensing Choices
Off-the-shelf industrial cameras with global shutter and GigE Vision or USB3 interfaces are the typical entry point; 5-megapixel sensors at 50-100 frames per second are sufficient for per-layer triggered capture on most FDM and PBF systems, per the DTU vision-pipeline design [S4].
For thermal overlay, mid-wave infrared cameras sensitive in the 3-5 µm band are used to capture melt-pool and cooling-region temperatures, then registered to the visible-light layer image by a homography transform calibrated to the build plate [S8].
On-machine illumination must be diffused and constant; specular highlights on metal powder surfaces saturate the image and break segmentation, so ring lights or coaxial LED arrays are the typical fielded solution rather than point sources [S4].
Software stacks lean on PyTorch or TensorFlow for the deep-learning classifier, with OpenCV handling the calibration, segmentation, and CAD-to-image warping steps, and a data-acquisition layer that timestamps each image to the build job's layer counter rather than wall-clock time [S1][S3].
Standards, Sourcing, and Traceability
No single ISO or ASTM standard governs layer-image-based defect detection methodology in 2026; quality-system requirements still flow from ISO/ASTM 52900 (AM process categorisation) and ISO/ASTM 52925 (specification of AM parts), while the layer-image sensor itself behaves like a process-control instrument whose calibration is documented per the OEM's quality plan [S1].
A 2026 Penn State MCREU study combined infrared thermography with computer vision and machine learning, registering thermal and visible images frame-by-frame so the defect decision is traceable to a specific layer number and timestamp in the build log [S8].
For sourcing decisions, the 2023 PMC survey recommends prioritising systems that publish their training-dataset provenance, defect taxonomy, and per-process detection accuracy, since proprietary black-box monitors cannot be re-validated on a new part geometry [S1]. Engineers specifying in-line vision should also confirm that the system can export raw layer images, not just pass/fail flags, so that warranty claims and root-cause analysis remain tractable after the build completes [S1][S2].
Selection Criteria for Engineers Specifying a Layer-Imaging System
For FDM or material-extrusion lines, the minimum viable system is one calibrated camera, one diffuse light source, and a CAD-to-image comparison pipeline, which the DTU work showed is sufficient to catch warpage, delamination, and extrusion failure in real time [S4].
For metal PBF, the system must include either a thermal camera, a melt-pool photodiode, or a coaxial visible-light camera inside the chamber, since the failure modes (lack of fusion, balling, recoat streaks) show up as thermal or texture anomalies that pure geometric comparison misses [S5][S8].
For mixed-vendor shops running both polymer and metal processes, a vision-imaging front end with a swappable classifier backend is preferable, so the same hardware can be redeployed across vision imaging and thermal imaging camera workflows without rewriting the data-acquisition layer.
Engineers should also weigh integration cost against the cost of scrapped builds: layer-imaging systems pay back fastest on long-run, high-value parts where a single failed build represents more than the cost of the vision system, and pay back slowest on short-run prototypes where manual inspection is acceptable [S1].
Process-Control Integration and Data Pipeline
Layer images should be written to a time-series database keyed by build job ID and layer number, so that downstream quality analytics can correlate defect flags with process parameters (laser power, scan speed, chamber pressure, feed rate) stored in the same record [S1][S2].
Most 2023-2025 implementations use a simple threshold on the pixel-deviation histogram to fire a stop-build signal when a layer's deviation exceeds a process-specific baseline; newer 2025-2026 work replaces the fixed threshold with a learned anomaly score from an autoencoder or one-class SVM trained on nominal-build layer images [S3][S6].
The data pipeline is also where the additive-manufacturing process-control stack intersects with broader factory analytics, since layer images can be treated as a sensor stream alongside pressure transmitter and flow meter data when correlating build defects to upstream gas-flow or chamber-pressure excursions.
A practical point often missed: the camera must be re-calibrated whenever the build plate is removed or the chamber is opened for service, because the camera-to-plate transform is the geometric anchor of the entire CAD-to-image comparison [S4].
Failure Modes the Vision Stack Itself Can Introduce
Spurious stops from lighting drift are the most common field issue: a ring LED that dims 10 percent over a year of service will shift the segmentation threshold and cause false-positive defect flags on nominally good layers [S4].
Specular saturation on metal powder, particularly on aluminium and titanium alloys, can mask recoat-streak defects and let them pass through undetected; the fix is polarised illumination and a cross-polarised camera filter pair [S8].
Camera-window fouling from condensation or stray powder is a chronic issue in PBF chambers; the 2025 high-precision monitoring review recommends scheduled in-situ window cleaning or a sacrificial glass slide that is changed between builds, since the alternative is a slow drift in measured layer geometry that mimics part-scale warpage [S2].
Finally, the data-storage cost of full per-layer image stacks on a multi-day build can run into terabytes per build job, so the pipeline should include on-the-fly compression and a retention policy that keeps raw images only for flagged layers, with downsampled thumbnails retained for the rest [S1][S2].
For Sourcing and Specification Engineers
Match the imaging stack to the dominant failure mode, not the marketing brochure. Polymer FDM: one RGB camera plus CAD-to-image comparison is the cost-optimised baseline. Metal PBF: add a thermal overlay or melt-pool photodiode to catch subsurface precursors. Binder jetting and directed energy deposition: confirm the vendor has published detection-accuracy data on your specific process before specifying [S1][S4][S5][S8].
Request raw-image export, calibration certificates, and a defect taxonomy that maps to your part-quality acceptance criteria; a system that only outputs pass/fail cannot support warranty or root-cause analysis when a flagged part is contested [S1][S2].
Trackable signals to watch: ASTM F42 committee activity on in-process monitoring metrics, any revision of ISO/ASTM 52925 to include image-based inspection clauses, and published benchmark datasets from NIST or the ASTM AM CoE that allow cross-vendor accuracy comparison. A practical near-term option is reviewing how a 2D vs 3D vision for robot guidance decision matrix applies to AM layer imaging, since the same trade-offs in field of view, depth resolution, and integration cost carry over.