Abstract:
A printed circuit board (PCB) is an essential part of every electronic equipment.
Because of technical innovation and the rise of the consumer market, the requirement
to make PCBs in large quantities has become a need throughout the years. Small flaws
in a PCB signal path can have a negative impact on an electrical device's operation.
Initially, the PCB inspection quality control procedure was done manually, which was
not only time-consuming but also prone to human errors. Later research focused on
traditional machine vision techniques such as template matching, image subtraction,
and morphological operations. These approaches were efficient and timesaving, but
they were susceptible to noise, orientation, and size. Faults in complicated patterns
were difficult to identify, and only a limited number of defects could be recognized
using traditional machine vision. To address these concerns, we propose a PCB fault
detection system based on the most recent version of you-only-look-once (YOLOv5).
YOLOv5 weights pre-trained on the Common Objects in Context (COCO) dataset
were used to execute transfer learning on the bare PCB defect dataset, which consisted
of 693 pictures with 6 different types of faults. Furthermore, data augmentation was
performed to improve the model's training performance and robustness. The medium
YOLOv5 model produced promising test results with a mean average precision (mAP)
of 96.92 percent, precision of 97.18 percent, and recall of 96.86 percent.
