A global FMCG giant needed more than traditional QA. As their product lines scaled, minor surface damage and dimensional inconsistencies began slipping through legacy systems. We built an edge-first inspection platform that uses 3D vision and smart damage detection to spot flaws in real time, even on rapidly changing packaging formats. The result: faster lines, fewer errors, and QA that scales with demand.
The challenge
A global Fast-Moving Consumer Goods enterprise shipping thousands of packaged products hourly struggled maintaining consistent quality across expanding box sizes and configurations. Traditional vision-based QA caught obvious defects, but frequent format changes, fast-moving lines, and variable packaging materials allowed subtle dimensional errors and minor surface damage to slip through undetected.
Core problems:
- Conventional off-site data processing risked slowing production lines or causing delays from network connectivity issues.
- Operators faced manual efforts updating manifests for each box type, introducing error and inventory mismatch risks.
- Rigid inspection thresholds designed for single SKU lines required recalibration whenever introducing new SKUs.
The customer needed precise, real-time shape measurement and surface damage detection - all while fitting smoothly into established manufacturing execution software and supporting multiple SKUs on single conveyor lanes.
Our approach
The solution combined a full edge-based approach with 3D sensors for dimensional verification and RGB cameras for surface inspection.
Real-time scanning
Dimension estimation using high-resolution point clouds detected minuscule deviations in length, width, and height while switching between different box sizes. Multiple-angle photography identified tears, dents, label misprints, scuffs, and other damage. Automated reference checks on known standards minimized manual intervention and calibration.
Edge processing
All scanning and image analysis occurred locally, reducing latency and eliminating network dependency. Packages flagged as out-of-spec triggered automated sorting lines, separating defective or dimensionally incorrect items before downstream processing.
Smooth integration
API-based interfaces updated facility software with exact package dimensions and QA results. The solution fit onto existing conveyors without full-scale re-engineering, keeping disruption to the production floor minimal.
Inspection architecture
High-fidelity 3D cameras
Each station’s 3D cameras captured dense point clouds across varying conveyor heights, speeds, and packaging materials. Internal algorithms filtered movement noise in real time, producing clean geometry data even at line speeds exceeding 100 units per minute.
Lightweight AI models
The damage detection pipeline and models were optimized for edge-grade processors, delivering real-time results without sacrificing accuracy. Model inference ran under 50ms per frame, leaving headroom for the parallel dimensional checks on the same hardware.
Automated updates
The system learned new baseline dimensions when introducing box styles. Operators confirmed once, then software continuously adapted - eliminating the recalibration downtime that plagued the previous setup.
Intuitive dashboards
Operators viewed pass/fail rates and live feeds on a centralized interface. Unusual spikes - such as a sudden uptick in corner tears on a specific line - triggered immediate review flags, letting supervisors intervene before defective batches accumulated.
Technical approach
Two parallel inspection pipelines ran side by side: surface integrity (local checks) and dimensional accuracy (global checks).
Sensor layer
Each conveyor station housed two sensors - one 3D scanner overhead and multiple angled RGB cameras - with an industrial-grade microcontroller handling data intake. The multi-angle setup caught damage that a single viewpoint would miss, such as underside scuffs or side-panel dents.
Edge compute node
A compact, specialized processor ran two parallel QA modules: local patch analysis for damage detection and global dimension analysis for shape verification. Each module processed data locally, generating pass/fail flags without cloud round-trips. This kept latency under 100ms from scan to decision.
Central orchestrator
A supervisory software layer connected each edge node to the facility’s manufacturing execution system (MES). It aggregated data for real-time dashboards and batch reports, giving plant managers a single view of quality across all lines without polling individual stations.
Local damage detector
The damage detector split incoming images into localized patches, inspecting each against known “normal” references via learned models. Each patch received a defect score indicating tear, scuff, or mislabel likelihood. By scoring at the patch level, the system pinpointed where on a package the damage occurred, not just whether damage existed.
Global dimension verifier
Outlier rejection algorithms pre-filtered spurious 3D points, then the verifier computed estimated package dimensions (length, width, height) from the cleaned point clouds. Results were compared against SKU specification tolerances, flagging packages that fell outside acceptable bounds even by a few millimetres.
MES integration
Edge nodes posted results - damage scores, dimension data, pass/fail flags - to message queues or REST endpoints monitored by MES. Failed packages triggered mechanical diverters removing items from main flow for manual review. Staff accessed live dashboards showing defect percentages, common damage types, dimension compliance rates, and line slowdowns.
The impact
Reduced downtime and rework
Instant detection of dimension or surface issues prevented entire batches of incorrect packaging from traveling further down the line. That cut rework times by over 40% compared to the previous system, which only caught major anomalies. The faster feedback loop also meant operators could identify root causes - a misaligned conveyor guide, a faulty sealer - before those issues compounded.
Improved accuracy across SKUs
By automatically tuning sensors to new box sizes and design features, the solution maintained near 100% detection accuracy. Even subtle shape differences triggered an alert before merging into the final shipping queue. This held true across the full range of SKUs running on each line, including seasonal and promotional packaging variants.
Simplified manifest management
The MES received package dimensions and QA status in real time, eliminating manual data entry. Operators no longer spent time logging box measurements or reconciling inspection records. That removed frequent human error sources and saved hours of administrative work each shift.
Future growth
Since the platform was deployed on the edge - using industrial-grade sensors with onboard compute - adding new lines or expanding to other facilities is straightforward. Each inspection node runs autonomously, yet stays synchronized with the central database. The architecture supports scaling without rearchitecting.
Lessons learned
Running inspection on the edge removed the bottleneck of external servers and network latency. For high-speed packaging lines where milliseconds matter, processing data at the sensor proved to be the single most effective architectural decision we made.
Static detection thresholds break down quickly in FMCG, where packaging changes are constant and SKUs rotate frequently. A learning-based approach - where the system adapts its thresholds as new packaging formats appear - kept misclassification rates low without requiring manual recalibration.
Catching defects early in the line, before packages mixed with conforming stock, reduced re-sorting and disposal costs substantially. The earlier a defective item exits the flow, the less labor and material gets wasted downstream.
Conclusion
For this large-scale FMCG client, the RGB-D-based quality assessment solution did more than improve QA metrics - it changed how they managed inventory and assured product quality. By unifying shape verification and surface analysis within a single, automated workflow, the plant could accommodate new product sizes while achieving near-perfect detection rates for damaged or mislabeled packages. Everything tied back into their existing manufacturing systems, keeping disruption low and ROI high.
This project demonstrates Algorithmic’s computer vision applied at industrial scale - real-time detection, edge deployment, and integration with existing manufacturing systems. The backend infrastructure was designed for autonomous edge operation with centralized monitoring, a pattern we apply across factory, logistics, and quality assurance environments.
