DeepVision Accuracy Methodology

Indus Vision claims 97–99.5% defect detection accuracy with <1% false positives for DeepVision deployments. This page explains how those numbers are measured and what they mean.

How We Measure Accuracy

Accuracy figures are derived from production deployment data collected across customer sites, not from controlled lab benchmarks. Each deployment undergoes a structured validation process:

1. Baseline study — We collect a representative sample of at least 5,000 inspected units, including manually verified known-good and known-defective parts, before DeepVision goes live.
2. Parallel run — DeepVision runs alongside existing manual inspection for a minimum of 2 weeks. Both decisions are recorded independently.
3. Ground truth reconciliation — A sample of DeepVision decisions (including marginal cases) is reviewed by trained quality engineers to establish ground truth labels.
4. Metrics calculation — Detection rate (true positive rate), false positive rate, and overall accuracy are calculated from the reconciled dataset.

Definition of Terms

• Detection rate (sensitivity) — percentage of actual defective units correctly identified as defective by DeepVision. Range: 97–99.5% depending on defect type and industry.
• False positive rate — percentage of non-defective units incorrectly flagged as defective. Target and achieved: <1%.
• Accuracy range explanation — the 97–99.5% range reflects variation across defect types. Surface scratches on flat steel achieve 99.5%. Complex multi-class defects on textured substrates achieve 97%. All figures are from real deployments.

Industry-Specific Notes

• Automotive stamping — validated on crack, burr and dimensional defect classes across ferrous and non-ferrous materials.
• Pharmaceutical tablets — validated on chip, crack, colour deviation and print defect classes at speeds up to 200,000 tablets/hour.
• FMCG packaging — validated on label, seal and fill-level defect classes at conveyor speeds up to 600 units/minute.
• Electronics PCB — validated on solder, component presence and orientation defect classes across SMT and through-hole assemblies.

Continuous Model Improvement

DeepVision models are retrained periodically using new production data from each deployment. Accuracy figures are re-validated after each major model update. Customers receive updated accuracy reports on request.

Request a Validation Report

Prospective customers can request a detailed validation report for their specific industry and defect types. Contact us →