The Inspection Challenge
Counting cigarette buds by hand means physically touching each roll — direct contact that risks contaminating the product and can deform the roll or its filter end. The result is a high-volume task that is both unsafe for product handling and unrealistic at line pace.
In regulated tobacco packaging, a count that is not tied to a specific tray, SKU, and timestamp is a traceability gap as much as a quality one. Reworking a misfilled batch after packing costs far more than catching it at the tray stage.
| Risk | Business Impact |
|---|---|
| Physical handling risk | Manual counting means touching each roll — contamination and roll/filter deformation risk affecting product functionality |
| Assumed, not verified, counts | Every tray ships on a nominal fill assumption, not a measured one |
| Under/overfilled cartons downstream | Pack-out errors surface only after material is already boxed |
| No batch-level audit trail | Manual counts cannot be traced back to a specific tray or shift |
| Regulatory exposure | Track-and-trace mandates expect verifiable, logged identification per batch |
Why Traditional Inspection Falls Short
Manual counting is not really an option for a ~4,000-item tray. It requires direct handling, degrades under fatigue, and provides no image-backed evidence trail.
| Limitation | Why It Matters Here |
|---|---|
| Direct handling required | Needs to stay sanitary and be handled delicately — impractical to guarantee by hand |
| Speed vs. volume | A ~4,000-item tray is not practically hand-countable at line pace |
| Fatigue-driven drift | Accuracy degrades measurably within the first 20–30 minutes of a shift |
| No data trail | A verbal or written count cannot be traced back to an image or a specific tray |
The Machine Vision Approach
The proposed approach is a single inspection station that captures the full tray in one image, identifies the tray/SKU, and counts the buds within a target cycle time. Every cycle produces an image-backed result that can be reviewed and audited later.
The system uses a high-resolution camera over the tray with diffuse illumination, a compact smart camera for barcode/OCR identification, and a deep-learning model that detects and counts individual buds directly rather than relying on a fixed grid or template.
Expected Outcomes & ROI
| Metric | Before → After |
|---|---|
| Count verification | Assumed → measured every tray, ±50 tolerance |
| Cycle time per tray | Not measured / bottleneck → ~6 sec target, single trigger-to-result |
| Audit trail | None → full image + count + model ID + timestamp per tray |
| SKU / model linkage | Manual, error-prone → automatic, decoded at point of count |
| Scalability across SKUs | New SKU adds manual variability → same model applies across SKUs, given uniform bud geometry |
Implementation Considerations
Because cigarette-bud geometry is highly uniform across SKUs, the same trained model can generalize across the portfolio with relatively few sample images. The rollout scales by deployment footprint: pilot, integrate the data, then replicate.
