深度学习
计算机科学
人工智能
印刷电路板
现场可编程门阵列
推论
目标检测
帧速率
计算机视觉
帧(网络)
计算机硬件
虚假关系
对象(语法)
实时计算
嵌入式系统
模式识别(心理学)
图像处理
还原(数学)
软件部署
集成电路
图像(数学)
电路设计
机器视觉
信号处理
特征提取
自动光学检测
作者
Lee Jia Qian,Nasri Sulaiman
出处
期刊:
[Institute of Electrical and Electronics Engineers]
日期:2025-11-24
卷期号:: 137-141
被引量:1
标识
DOI:10.1109/icsys65149.2025.11379920
摘要
This paper presents the development and implementation of a real-time Printed Circuit Board (PCB) defect detection system using deep learning on FPGA. A YOLOv3-Tiny object detection model was trained using a publicly available dataset containing 9014 high-resolution images annotated with six common PCB defect types: missing hole, mouse bite, open circuit, short circuit, spur, and spurious copper. The training was conducted using the Hank AI Darknet framework, and the best weights were achieved at iteration 28,819 with a mean Average Precision (mAP) of 99.55%. To enable deployment on resource-limited hardware, the trained model was optimized using the DNNDK toolchain. The model was quantized to 8-bit fixed-point and compiled for execution on the Deep Processing Unit (DPU) integrated into the Xilinx PYNQ-Z2 FPGA board. Both static image inference and real-time inference were successfully executed. The FPGA-based system achieved a real-time frame rate of 2.015 FPS and a post-deployment mAP of 28.17% at IoU threshold of 0.5. Despite a reduction in detection accuracy due to quantization, the system remains suitable for a standalone, power-efficient, and cost-effective PCB defect detection system for real-time industrial applications.
科研通智能强力驱动
Strongly Powered by AbleSci AI