Solution Brief · AMD Kria™ SOM Heterogeneous Compute

Wire-Speed AI Inference for High-Speed Inspection

YantraVision combines high-bandwidth line-scan imaging, FPGA acceleration, and AI inference to detect defects down to 1 mm on production lines running at 20 m/s. Built on the AMD Kria™ K26 SOM, the system processes every acquired line with less than 2 ms detection latency while distinguishing true defects and contaminants from natural material variation.

The Evolution

From deterministic rules to AI-assisted decisions

Since 2016, YantraVision has designed and deployed AMD-based machine-vision inspection and sorting systems for industrial lines that cannot slow down: cotton, grain, recycled fiber, pharmaceutical print, and paper. Across these applications, deterministic rules remain effective when defects are visually distinct and process conditions are stable. As inspection targets become smaller and natural material variation increases, however, rules alone cannot always separate a true contaminant or defect from an acceptable variation. YantraVision is addressing that gap by adding AI-based inference to its high-speed imaging and FPGA processing platform.

Paper & Print Cotton Grain Recycled Fiber Pharmaceutical Print
The Twin Challenge

See small defects at speed. Decide what they mean.

The installed YantraVision system inspects a paper web running at up to 20 m/s, with end-to-end detection latency of less than 2 ms. In the installed system, it detects defects down to 1 mm, including oil smears, oil splashes, pinholes, contamination flecks, creases, and edge tears. Rather than ejecting material in real time, the system records each detected defect's machine- and cross-web coordinates in a defect roll map, enabling the downstream die cutter to locate, remove, and patch the affected section.

Live capture from YantraVision's inspection software showing the paper web feed and defect information panel
Live capture from YantraVision's installed paper-web inspection system.

The application presents two tightly coupled engineering problems. First, the imaging system must capture enough spatial detail to reveal small defects while the web moves at production speed. That creates a high-bandwidth data stream that must be acquired and processed continuously.

The deployed capture path uses a 4K line-scan camera in Camera Link Full mode at an 85 MHz clock and 42,000 lines/s. At 4,096 pixels per line and 8-bit RGB, it carries approximately 516 MB/s of pixel data — about 76% of the interface's 680 MB/s nominal capacity — while retaining headroom for continuous operation.

Second, acquiring the image is only half the problem. Natural variations in the paper can overlap with defects in size, color, intensity, or texture, making fixed thresholds prone to misses and false positives. AI inference provides a learned decision boundary, but it is also compute-intensive. YantraVision therefore had to solve both constraints together: sustain the image bandwidth and complete inference within the available time budget.

Paper rolls moving through a production line
Paper roll stock inspected on the production line.
Synthetic paper web sample with unmarked defects
Illustrative reference: synthetic paper-web sample at realistic scale. Defects are intentionally unmarked here; the palette below identifies each type and its typical size.
Defect palette with representative defect types and a minimum detected defect size of 1 mm
Defect palette: representative defect types detected by the production system, shown enlarged for clarity.
Solution

Extending a rules engine with AI inference

YantraVision built the Falcon Compute platform to perform deterministic, low-latency image processing in FPGA fabric at line speed. Its rules engine continues to provide fast, repeatable detection for well-defined conditions. The AI upgrade addresses the harder cases, where defects overlap with normal web variation and cannot be separated consistently by fixed thresholds alone.

Comparison of fixed and learned decision boundaries using the same production data.

Decision-boundary comparison showing the rule-based threshold strategy versus the AI inference learned-boundary strategy across overlapping defect types and the accepted class
Table comparing rule-based detection against AI inference detection across three defect types plus one acceptable anomaly, with size and intensity as controls

YantraVision deployed the upgraded inspection pipeline on the AMD Kria™ K26 SOM, based on the Zynq™ UltraScale+™ MPSoC architecture. FPGA logic handles high-throughput acquisition and feature extraction, while the processing system performs AI-based classification. This heterogeneous design evaluates every acquired tile continuously, without frame skipping. The learned classifier distinguishes defects from normal variation and identifies the defect type, enabling the appropriate downstream action instead of treating every anomaly alike.

YantraVision inspection architecture showing continuous line-scan processing, FPGA feature extraction, AI classification on AMD Kria K26 SOM, and defect roll mapping
System architecture: FPGA-based feature extraction feeds real-time classification and decision-making on the same AMD Kria™ K26 SOM, continuously at the stated line speed. On this paper line, every decision is written to the defect roll map and passed to the downstream die cutter; no pneumatic ejector is used.
Installed inspection hardware on this line
Line scan camera used on the paper inspection line
Line Scan Camera
High-speed, high-resolution imaging of the moving paper web.
AMD Xilinx Kria K26 SOM mounted on the YantraVision carrier board
Kria™ SOM + YV Carrier Board
AMD Kria K26 SOM on YantraVision's carrier board with Camera Link connectivity.
YantraVision Vision Processing Module (Falcon Compute) enclosure
Falcon Compute Enclosure
YantraVision's Falcon Compute enclosure, which houses the Kria SOM.
High-intensity LED lighting unit used for illumination
LED Lighting Unit
High-intensity illumination engineered to reveal defects across the web.
Hardware from the installed YantraVision system.
Results

Proven performance

20 m/s
Sustained production line speed
<2 ms
End-to-end detection latency
1 mm
Minimum detected defect size
Continuous
Every acquired tile evaluated

In the installed system, the complete imaging and inference path operates at line speed on the AMD Kria platform. Each detected defect is recorded in the roll map at its machine- and cross-web coordinates, enabling the downstream die cutter to act on the correct section of material. This supports continuous inspection, defect classification, and defect-specific downstream handling.

At 20 m/s and 42,000 lines/s, the machine-direction sampling interval is approximately 476 µm per line, providing just over two line samples across a 1 mm feature. Actual detectability also depends on contrast, optics, illumination, exposure, and feature orientation.

Explore More

More from YantraVision

AI Acceleration →
FPGA pipelines, AMD Ryzen AI NPU inference, and GPU analytics for real-time industrial vision.
AI Model Porting →
Moving trained models from GPU and cloud to FPGA, AMD NPU, and edge runtimes with validated accuracy.
Custom OEM Solutions →
Full-stack vision engineering for machine builders, from sensor to decision.
Falcon Sorting Platform →
Three-layer FPGA sorting intelligence with sub-2 ms latency for OEM machine builders.
AMD ECS 2026 Recap →
Our two live AMD Ryzen AI NPU demos from the Embedded Computing Summit in Bengaluru.
Case Studies →
Silicon validation, FPGA vision pipelines, and inspection system deployments.

Adapt this architecture to your inspection platform

For OEMs developing inspection equipment, YantraVision can adapt the platform to new materials, defect classes, imaging configurations, and production speeds, while integrating application-specific cameras, lighting, interfaces, and downstream controls.

Discuss Your ApplicationExplore YantraVision