AI-Powered PCB Inspection: Defect Detection in Real Time with Microchip PolarFire® SoC FPGA and SpanIdea AI Technology

From Human Imperfection to Artificial Intelligence (AI) Precision

Imagine a master watchmaker inspecting a luxury timepiece with hundreds of identical, tiny gears—each critical to the watch’s function. Though the gears are visible to the naked eye, the sheer repetition of checking everyone for defects or misalignment strains even the steadiest hands and sharpest eyes. Now, picture an AI-powered inspection system designed for watches. It can analyze intricate components at high speed, detect microscopic imperfections and ensure precise alignment—far beyond human capabilities. Similarly, an AI-powered Printed Circuit Board (PCB) inspection system scans each board in seconds, accurately identifying missing screws and defects with bounding boxes—eliminating the need for human expertise or visual inspection.

This is the breakthrough that we achieved by deploying an AI application on Microchip PolarFire® SoC Video Kit. By extracting patches from high-resolution PCB images, classifying these patches as “screw present” or “screw missing,” and overlaying the results directly onto the PCB image, we have revolutionized PCB quality control—eliminating manual effort and enabling a seamless, fully automated inspection process.

By reading this blog post, you will gain insights into:

• How Spanidea’s AI-powered image processing simplifies PCB quality inspection by detecting missing screws
• Why Microchip PolarFire SoC Video Kit is a game-changer for low-power, real-time AI applications
• Real-world performance metrics that demonstrate speed, accuracy and efficiency of defect detection in high-volume production environments

The Critical Need for Automated PCB Inspection

In modern electronics, flawless PCB assembly is essential, as a single missing component can lead to critical failures in aerospace, automotive and medical devices. Traditional inspection methods rely on human inspectors, who are susceptible to fatigue and inconsistencies. Edge AI addresses these challenges by enabling real-time, adaptive analysis directly on the factory floor. With low-power FPGA fabric and RISC-V cores the Microchip PolarFire SoC is optimally designed for deploying vision models directly on production lines, eliminating the need for cloud-based processing.

The Technical Breakdown

Training the Classifier – From Data to Deployment

In Industrial automation, detecting subtle defects demands both precision and speed. To meet this challenge, Spanidea developed a custom convolutional neural network (CNN) trained on proprietary datasets using TensorFlow. The model demonstrated robust accuracy in identifying missing screws during critical inspections, establishing the foundation for reliable automated quality control. To optimize the FP32-trained model for embedded deployment, we streamlined its footprint by quantizing it to 8-bit integer. This process involved converting the TensorFlow graph into OpenVINO’s Intermediate Representation (IR) and leveraging Microchip’s VectorBlox™ Accelerator SDK for hardware-aware quantization, a key step that slashed computational costs without compromising accuracy. The final output, a compact ‘*. vnnx blob’, was specifically optimized for FPGA execution, ensuring seamless operation in resource-constrained environments. The overall flow of the process is described in Fig.1 below. The input images provided by the camera module of the board need to be transformed to get ingested by the model deployed on the board. The scripts provided in VectorBlox SDK were modified to enable this interfacing between raw images and the deployed model’s input. Post-processing functions translated raw probabilities into actionable insights, enabling real-time defect detection and seamless integration into production workflows.

The deployment pipeline designed for the PolarFire SoC FPGA Video Kit leverages its onboard RISC-V processor to embed AI inference directly into the FPGA fabric. By eliminating the need for external GPUs, the system achieved real-time performance with minimal latency and power draw—critical for uninterrupted industrial operations. FPGA-specific optimizations, enabled by the VectorBlox SDK, ensured efficient utilization of hardware resources, from memory bandwidth to parallel compute units. This approach not only simplified integration but also preserved the model’s diagnostic accuracy, proving that edge-ready AI can thrive in low-power, high-reliability settings. The result? A scalable, energy-efficient solution that marries cutting-edge deep learning with the pragmatic demands of industrial automation.

Image Cropping – Extracting Screw Patches

We are given the raw PCB image as shown in Fig. 2. To crop the screw from the PCB, the flow leverages Microchip’s Image Scaler IP. Since the location of screws is known as a priory, this IP is modified to precisely crop and resize screw regions from high-resolution PCB images. The modified IP resizes screw images to 128x128 pixels, aligning with the input requirements of the deployed custom CNN model. These images are then ingested by the model for detection. Fig. 3 shows different classes which the model will classify.

Fig. 2. DUT (Device Under Test) image

Fig. 3. Datasets Classes (a) Screws (b) Missing Screw (c) Background

Overlaying Results – Bounding Boxes in Real Time

After classification, the PolarFire SoC’s RISC-V core maps result back to the original PCB image coordinates using pre-optimized C/C++ scripts from the VectorBlox SDK. Bounding boxes and labels are overlaid onto the live video feed in real-time as shown in Fig. 4, enabling instant visual feedback for operators.

Fig. 4. Final result: Missing Screws

Performance Metrics: Optimized for Speed and Scalability

• Low Latency: Processes each image in ~21 milliseconds efficiently handling cropping, classification and overlay.
• High Throughput: Handles 47 images per second, ensuring seamless scalability for high-volume production lines without bottlenecks.

Note: The term image above refers to patches cropped from the original PCB image.

Next Steps – Implement AI-Driven Application on PolarFire

To get started with deploying your AI application on the PolarFire SoC FPGA Smart Embedded System, explore the VectorBlox-SDK  and leverage the PolarFire SoC video kit for hardware testing. Harness the power of low-power AI inference with PolarFire SoC video kit and VectorBlox-SDK today!

At Spanidea, we specialize in delivering cutting-edge AI algorithm models tailored for a wide range of industrial applications, including PCB inspection, defect detection, predictive maintenance and smart automation. Our expertise in optimizing AI for FPGA and edge deployments ensures high-performance, scalable and energy-efficient solutions that drive innovation in industrial AI. Partner with us to accelerate your AI-driven transformation.