An Efficient Tiny Defect Detection Method for PCB With Improved YOLO Through a Compression Training Strategy

Tiny defect detection is a knotty task in industrial electronics production. Existing traditional and deep learning methods have achieved satisfactory performance, however, they still face challenges in accuracy, generalization ability, and computational complexity. Therefore, this study designs a t...

Full description

Saved in:
Bibliographic Details
Published in:IEEE transactions on instrumentation and measurement 2024, Vol.73, p.1-14
Main Authors: Zhou, Wen, Li, Changyi, Ye, Zhiwei, He, Qiyi, Ming, Zhe, Chen, Jingliang, Wan, Fang, Xiao, Zhenhua
Format: Article
Language:English
Subjects:
Accuracy
Circuit boards
Complexity
Compression training strategy
Computer networks
Datasets
Defects
Detectors
Feature extraction
Head
Image coding
Image compression
Image quality
Image resolution
Industrial electronics
printed circuit board (PCB)
Printed circuits
Real-time systems
tiny defect detection
Training
YOLO
you only look once (YOLO)
Citations: Items that this one cites
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Tiny defect detection is a knotty task in industrial electronics production. Existing traditional and deep learning methods have achieved satisfactory performance, however, they still face challenges in accuracy, generalization ability, and computational complexity. Therefore, this study designs a tiny defect detection-based you only look once (TDD-YOLO) model and proposes an innovative compression training strategy to train on low-resolution images and test on original images. First, a four-ME layers structure is adopted to the backbone network, to integrate more underlying information and extract effective features. In addition, a miniature detection head is incorporated into the head network to improve the accuracy and generalization performance of you only look once (YOLO). Meanwhile, TDD-YOLO introduces wise intersection over union (W-IoU) to reevaluate the loss of bounding box regression (BBR) and reduce false negatives by fitting the model well to regular quality anchor boxes. Finally, an image compression method at different ratios is applied in the proposed compression training strategy, to reduce computational complexity and surprisingly further improve accuracy. Comprehensive experiments on several variable compressed datasets which are based on a public printed circuit board (PCB) defect dataset validate the effectiveness of our theoretical approach and illustrate that our proposed method outperforms state-of-the-art methods.
ISSN:0018-9456
1557-9662
DOI:10.1109/TIM.2024.3390198