AI Platform Overview

Hyundai MIB International has developed a breakthrough in banknote counting and sorting by combining an Analog NPU (Mythic M1076) with an Advanced CIS sensor. This enables deep learning-based image analysis at full resolution — a first in the desktop banknote machine industry.

Hyundai MIB International co-developed this platform with Mythic and holds exclusive rights to use Mythic's analog NPU technology in the banknote handling machine industry. No other banknote machine manufacturer has access to this technology.

AI Platform architecture: 5-Mode CIS Sensor captures banknote under IR, UV, and visible light, sends data to Zynq UltraScale+ 1EG SoC, which connects via PCIe to Mythic M1076 NPU (26 TOPS) for classification (denomination, counterfeit, fitness, tape) and external display output

At a Glance

	Old Technology	New Technology
Pixels analyzed	0.01%	100%
Classification	Rule-based (counterfeit) + SVM (denomination) on CPU	ResNet on Analog NPU
CIS modes	3 (IR-R, IR-T, Visible)	5 (+ UV, additional IR)
Tape detection	Mechanical (indirect)	CIS imaging + deep learning (direct)
Processing power	Zynq 7010/7020	Zynq 1EG + Mythic M1076 (26 TOPS)
Speed	1,000–1,500 NPM	1,000–1,500 NPM (same speed, far better accuracy)

The Problem with Old Technology

Traditional banknote counting machines — including our previous products — relied on a combination of rule-based programming for counterfeit detection and SVM (Support Vector Machine) for denomination classification, all running on a low-power CPU. For counterfeit detection, the old system used two separate SVMs — one analyzing 19 × 6 = 114 features from specific areas and another analyzing 22 × 5 = 110 features across the full note area — where each feature was the average value of a rectangular region. Combined with hand-written rules, this approach sampled only a tiny fraction of the available image information.

The old CIS sensor was limited to 3 spectral modes (IR Reflection, IR Transmission, Visible Light), missing the UV and additional IR channels that reveal details invisible to older sensors.

Read the full analysis of old vs. new approach →

Our New Technology

Component	Old	New
Main SoC	Zynq 7010 / 7020	Zynq 1EG (with PCIe support)
AI Processor	None (rule-based + SVM on CPU)	Mythic M1076 Analog NPU (26 TOPS)
Connection	—	PCIe (high-bandwidth image delivery)
CIS Sensor	3-mode (IR-R, IR-T, Visible)	5-mode (+ UV, additional IR)

What This Means for You

Better Detection

Full-image AI analysis examines every pixel — and the relationships between them across multiple spectral layers — to detect tape, graffiti, counterfeits, and fitness defects that traditional sensors miss entirely.

Smaller & Quieter Machine

The new CIS with full-range UV eliminates separate UV point sensors, and AI-based imaging replaces the mechanical tape detector. Removing these components makes the machine more compact and significantly quieter.

No Programming Required

With a data-driven approach, new currencies and counterfeits can be handled by retraining the model with sample data — even a dealer can do it, without the manufacturer's involvement.

Future: Unseen Counterfeit Detection

By learning the features of genuine notes through an encoder + decoder architecture, the system could detect counterfeits it has never seen before — eliminating the need to update software each time a new counterfeit appears. Preliminary testing is underway.

Unseen counterfeit detection requires an encoder model, which the next generation of the Mythic NPU will support. Current capabilities focus on decoder-based classification using ResNet.

Coming Early 2027

First AI-Powered Product: 2-Pocket Fitness Sorter

We are developing our first product with the AI platform built in — a 2-pocket fitness sorter powered by the Mythic M1076 Analog NPU and 5-mode Advanced CIS. This will be the first desktop banknote machine in the industry with deep learning-based full-image analysis.

The new machine replaces the mechanical tape detector with AI-based imaging — making it quieter, smaller, and dramatically more accurate in tape and counterfeit detection.

The 2-pocket fitness sorter is the first product to receive this technology. We will progressively apply the AI platform to our entire product lineup — from value counters to multi-pocket sorters.

Early Tester Program

We are accepting a limited number of early testers — selected dealers who will get access to the AI-powered machine ahead of general availability and have the opportunity to introduce this product into their market first.

Why We Need Partners

Deep learning models are only as good as their training data. To build highly accurate models for each currency, we need large volumes of well-classified banknote data — genuine notes, counterfeits, and various fitness conditions — collected from real-world operations in each local market.

No single company can collect this data alone. That is why we are partnering with selected dealers in each market to collaboratively build the AI models for their currencies.

How It Works

Collect Data

Using our machine, the dealer collects and classifies banknote images from their local market — genuine, counterfeit, fit, unfit, taped, etc.

Build the Model

We train the AI model using the collected data. The more data, the better the model performs on that currency.

Early Market Access

Early testers are the first to bring the AI-powered machine to their market — with a model built specifically for their local currencies.

What Early Testers Get

First-mover advantage — introduce AI-powered banknote processing to your market before competitors
Custom-trained models — AI models optimized for the specific currencies and conditions in your market
Direct collaboration — work closely with our engineering team during development

Interested in becoming an early tester? We are selecting partners on a per-market basis. Contact us to discuss the opportunity.