Anti-counterfeit Label Verification

Overview

A packaging/security client asked us to validate whether computer vision can tell authentic label prints from counterfeits and, if feasible, deliver a production-ready model and a mobile evaluation path. We proposed a staged plan: feasibility → dataset → modelling → mobile, with a go/no-go at each gate to control risk and budget

At a glance

Industry

E-commerce

Goal

Provide a system for reliable authenticity checks, ideally on-device

Scope

Feasibility study, dataset build, ML models, mobile SDK validation

Core Business Challenge

• Is the signal there? If humans can’t reliably tell authentic vs. counterfeit, CV may be infeasible.

• Data reality: Build a robust dataset (varied lighting/cameras) with annotated micro-markers.

• Deployment fit: Decide between server-side inference and on-device models across heterogeneous phones.

Our Approach

Stage 1 - Feasibility (human baseline)

Inspect physical prints under varied lighting to confirm discriminative features and define a capture protocol (≥30 authentic and ≥30 counterfeit samples).

Exit: “CV is viable” + differentiator list.

Stage 2 - Dataset

Collect ≥700 authentic and ≥700 counterfeit prints; capture images per protocol (potentially multiple light settings & camera models); annotate markers.

Exit: training/validation set ready

Stage 3 - ML Model

Train models to locate markers and classify authentic vs counterfeit; explore single-image vs multi-illumination series; compare deep nets with clustering baselines; stand up MLOps and package server-ready models.

Exit: ML models tested + metrics

Why it's low-risk high-reward for the client

• Stage-gated: each step has a clear “continue/stop” decision to avoid sunk cost.

• Data-first: feasibility checks and capture protocol reduce the chance of chasing nonexistent signals.

• Deployment-aware: we test mobile constraints early (compute, operators, accuracy drift across phones).

Outcome

We run the full discovery stage and found some big obstacles that could compromise the project in later stages. As we informed the client of the results, they decident no to pursue, appreciating our honesty and strategic approach.