Refined Defect Detector With Deformable Transformer and Pyramid Feature Fusion for PCB Detection

Printed circuit board (PCB) defect detection is a part of the quality control process, which detects and identifies predamage in finished products. However, it is difficult to detect them due to small defects. To this end, this article designs a refined defect detector (RDTor) with deformable transformer and pyramid feature fusion to precisely capture small defects for achieving the defect detection of PCB, where RDTor is comprised of three parts: multihead nonlocal transformer (MNT) module, multiscale pyramid feature fusion (MPFF) module, and adaptive defect detection (ADD) ones. Specifically, an MNT module is first developed to adaptively focus on the defect areas for highlighting defect features and suppressing nondefect background ones, an MPFF module is proposed to preserve the features of small defects much more as network deepening, and an ADD module is presented to adaptively perform the defect inspection for obtaining defect categories and defect prediction box. The experimental results on a large-scale PCB image dataset acquired from real-world industrial products show that the proposed method achieves the state-of-the-art accuracy in industrial applications, where it can achieve 99.6% (accuracy), 97.2% (precision), 98.8% (recall), 98.0% (F1-score), and 99.3% mean Average Precision (mAP) in terms of multiscale defects classification and detection results.

I. Introduction

Printed circuit board (PCB) has a wide range of applications in industrial products. Its quality directly affects the performance of electronic devices, so it is very important to identify and screen anomaly PCBs [1]. Conventionally, manual detection is usually performed to distinguish whether the defect is true or pseudo, which makes manual detection inefficient and costly, thereby making it difficult to meet PCB batch production requirements. Therefore, it is necessary to develop an automated PCB defects detection system to fast and accurately detect defects.