APMLF-YOLO: a progressive multi-level feature fusion algorithm for industrial PCB tiny defect detection

The Printed Circuit Board (PCB) relies on manual and electrical detection methods for quality inspection in its early stages. Traditional methods not only fail to meet the requirements of modern industry for detection efficiency and accuracy, but also make it difficult to achieve industrial deployme...

Full description

Saved in:
Bibliographic Details
Published in:Signal, image and video processing Vol. 19; no. 14; p. 1195
Main Authors: Wang, Shoubin, Li, Kun, Peng, Guili, Gao, Pengcheng, Xing, Zhihan, Zhang, Hanrui, Gao, Zimeng, Li, Youbing, Fang, Xinchang, Jing, Lewei
Format: Journal Article
Language:English
Published: London Springer London 01.12.2025
Springer Nature B.V
Subjects:
Accuracy
Algorithms
Automation
Circuit boards
Computer Imaging
Computer Science
Deep learning
Defects
Design
Efficiency
Feature extraction
Feature selection
Image Processing and Computer Vision
Industrial applications
Information technology
Machine learning
Manufacturing
Methods
Modules
Multimedia Information Systems
Object recognition
Original Paper
Parallel processing
Pattern Recognition and Graphics
Printed circuit boards
Printed circuits
Public safety
Redesign
Salience
Signal,Image and Speech Processing
Task complexity
Vision
Vision systems
Feature fusion
Scale transformation
High-precision detection of PCBs
Attention
Small targets
ISSN:1863-1703, 1863-1711
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:The Printed Circuit Board (PCB) relies on manual and electrical detection methods for quality inspection in its early stages. Traditional methods not only fail to meet the requirements of modern industry for detection efficiency and accuracy, but also make it difficult to achieve industrial deployment. Therefore, this paper proposes a high-precision defect detection model called the attention and multi-level feature fusion YOLO algorithm (APMLF-YOLO). The purpose of its design is to achieve high-efficiency and high-precision detection of PCBs. We redesign the Neck structure of the model and propose a multi-level feature fusion network module with multi-angle parallel processing. By integrating the MHSA mechanism into the AFPN structure and further adding a small object detection layer, the problem of multi-scale feature fusion is effectively addressed, enhancing the model’s perception of subtle defects. Subsequently, we designed a lightweight salient feature extraction module called CAM-Light to integrate into the backbone structure of the network, enabling the model to more effectively focus on defect areas. Finally, a sliding loss function is introduced to learn the importance of different input features. The method presented in this article demonstrates significant detection performance on the PCB dataset: The mAP50 achieves 99.1% with a 3.3% improvement, while mAP50-95 improved by 6.4%, the F1 score increased by 3%, and the number of model parameters is only 5.1 M. This model shows excellent robustness and comprehensive performance in small and complex defect detection tasks, making it more suitable for industrial applications.
Bibliography:ObjectType-Article-1
SourceType-Scholarly Journals-1
ObjectType-Feature-2
content type line 14
ISSN:1863-1703
1863-1711
DOI:10.1007/s11760-025-04738-9