.png){kind=icon}
.png){kind=icon}
Return
Attention-guided inaccurate supervised learning for conductive particle detection
.png){kind=icon}
PRE
AI
Z
H
J
H
G
T
DOI：10.1117/1.JEI.35.2.023042.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Deep learning methods have been widely applied in vision-based industrial detection, particularly in tasks such as conductive particle detection in the chip on glass and film on glass processes. However, existing methods heavily rely on supervised learning, requiring large amounts of accurately labeled data, which is impractical to obtain through manual annotation, leading to suboptimal detection accuracy. In this paper, we propose an attention guided inaccurate supervised learning approach to efficiently train a robust detection model with minimal manual intervention. First, a coarse dataset is constructed by utilizing target-specific features, such as rough localization of conductive particles via handcrafted feature extraction. Then, a task-specific network is developed to support the inaccurate supervision training scheme; we design a lightweight U-shaped network, UGA-Net, which incorporates a group attention mechanism to facilitate the inaccurate supervision training scheme and handle large image scale variations of conductive particle images. Finally, we propose the inaccurate supervision training scheme to iteratively refine the coarse dataset through confidence-based updates, enabling the model to learn from inaccurate labels and progressively improve its accuracy. Experimental results show that the proposed approach outperforms state-of-the-art methods in conductive particle detection while significantly reducing the need for manual intervention.
Keywords:
inaccurate supervision
light-weight neural network
UGA-Net
conductive particle
Journal
J
IF:
1
Papers:
109
Citations:
2.7K
Organization
Cited Papers