A real time segmentation network for lithium battery surface defect detection.

Area of Science:

• Materials Science
• Computer Vision
• Artificial Intelligence

Background:

• Lithium battery surface defect detection is crucial for industrial applications.
• Existing methods struggle with varying defect scales, especially tiny defects.
• High accuracy and real-time performance are essential for defect detection.

Purpose of the Study:

• To propose a novel network, Dual Attention Pyramid Segmentation Network (DAPSeg), for precise and real-time lithium battery surface defect segmentation.
• To address challenges of scale variation and the need for simultaneous high accuracy and speed.

Main Methods:

• Developed DAPSeg featuring a Selective Kernel Module (SKM) for adaptive multi-scale feature extraction.
• Employed a lightweight segmentation head with Blueprint Separable Layer (BSL) and Dual Attention Feature Fusion Module (DAFFM).
• Utilized a diffusion model for data augmentation on the LB-SD dataset to mitigate overfitting.

Main Results:

• DAPSeg achieved mIoU scores of 79.57% (LB-SD), 83.53% (MT), and 89.10% (MSD).
• The model demonstrated a processing speed of 74.09 FPS.
• Outperformed state-of-the-art models in balancing accuracy and inference speed.

Conclusions:

• DAPSeg offers a robust solution for lithium battery surface defect detection.
• The network achieves high precision and real-time processing capabilities.
• DAPSeg exhibits strong generalization performance across different datasets.