Packaging Inspection: Label Verification on GPU

The Challenge: 120 SKUs, 14 Lines, Two Recalls Last Year

A pharmaceutical contract packaging company near Cambridge operates 14 packaging lines processing over-the-counter medications, supplements, and medical devices for 35 client brands. With 120 active SKUs and frequent changeovers — some lines switch products four times per shift — mislabelling risk is constant. Last year, two mislabelling incidents resulted in product recalls costing a combined £340,000 in direct expenses (retrieval, destruction, investigation) and unmeasurable reputational damage with clients. One incident involved a children’s paracetamol carrying an adult dosage label; the other put the wrong allergen declaration on a supplement product. Both errors occurred during line changeovers when operators loaded the wrong label reel.

The existing camera-based verification system checks barcodes and batch codes but cannot verify label content — it confirms that a barcode scans correctly but not that the words on the label match the product inside the package. The company needs full label content verification running at line speed (200 units per minute).

AI Solution: OCR and Vision-Based Label Verification

A combined OCR and vision AI pipeline captures an image of every label at line speed, reads all text content (product name, dosage, ingredients, allergen warnings, batch code, expiry date), and compares it against the expected label specification for the current production run. The system detects: wrong product label, incorrect dosage information, missing allergen declarations, wrong language version, and batch/expiry code errors.

Running on a dedicated GPU server, the pipeline processes 200 labels per minute per line — 2,800 per minute across all 14 lines — with each label verified in under 300 milliseconds. Any mismatch triggers an immediate line stop and operator alert.

GPU Requirements

Label verification requires both OCR (text extraction) and NLP (content comparison against specifications). Processing 2,800 labels per minute demands high-throughput GPU inference with consistent low latency.

A single RTX 5090 handles all 14 lines with significant headroom. Private AI hosting ensures client product formulations and label specifications remain within GDPR-compliant UK infrastructure.

Recommended Stack

• PaddleOCR for high-speed text extraction from label images, optimised for small text at various orientations.
• YOLOv8 for label layout detection — verifying that all required fields (product name, dosage, warnings, barcodes) are present in their expected positions.
• Fuzzy string matching (RapidFuzz or FuzzyWuzzy) for comparing extracted text against the master label specification, accounting for minor OCR variations.
• GigE Vision cameras at each line station capturing labels at line speed.
• PLC integration via OPC UA for sending line-stop commands when mismatches are detected.

For generating batch documentation, add an LLM via vLLM. Use document AI to digitise client label specifications from PDFs and artwork files, automatically populating the verification database.

Cost Analysis

The two recalls last year cost £340,000 directly, with client relationship damage estimated at a further £200,000 in at-risk contracts. Eliminating mislabelling errors entirely through 100% label verification represents a minimum £340,000 annual risk reduction. The existing camera verification system costs £15,000 annually in maintenance and provides only barcode checking. The GPU-powered upgrade adds full content verification for a modest incremental server cost.

Beyond recall prevention, the system provides a complete digital audit trail of every label verified on every unit — invaluable during MHRA inspections and client quality audits.

Getting Started

Compile master label specifications for your 20 highest-volume SKUs, including all text fields that must be verified. Capture 1,000 label images from each line camera, covering different label orientations and lighting conditions. Train and validate the OCR pipeline, targeting 99.9% text extraction accuracy before deploying to production. Start on one line and expand after demonstrating zero false negatives over a 30-day validation period.

GigaGPU provides UK-based dedicated GPU servers for pharmaceutical and packaging workloads. Add an AI chatbot for operator access to packaging specifications and changeover procedures.