Advances in Automated Optical Inspection Using Attention-based Convolutional Neural Networks

Strategic Insights -

Researchers at Chungbuk National University have proposed an attention-based convolutional neural network (A-CNN) model for automated optical inspection (AOI) of small objects. The A-CNN model uses an attention module to adaptively extract regions of interest centered on small objects, improving classification accuracy. Experimental results show that the proposed A-CNN model achieves a classification accuracy of 99.92% and an inference speed of 33 frames/s on the NVIDIA Jetson Nano platform.

Key Takeaways:

  • The A-CNN model integrates an attention module to adaptively extract regions of interest centered on small objects, increasing the object-to-image area ratio and improving classification accuracy.
  • The A-CNN model achieves a classification accuracy of 99.92% and an inference speed of 33 frames/s on the NVIDIA Jetson Nano platform.
  • The proposed model outperforms state-of-the-art object detection algorithms, including YOLOv5, YOLOv7, YOLOv8, YOLOv9, and YOLOv10, in terms of both accuracy and latency.
  • The A-CNN model has a faster inference speed than the fastest YOLO model, providing $2.0\times$ speedup.
  • The A-CNN model is a practical solution for small object classification, with potential applications in real-time inspection and quality control.
  • The researchers propose that their model can be effectively trained end-to-end with minimal data labeling compared with object detection methods.

Statistics:

  • Classification accuracy: 99.92%
  • Inference speed: 33 frames/s
  • Object-to-image area ratio (OAR): Improved by using the attention module
  • Training time: Minimal data labeling required
  • Inference speedup: $2.0\times$ faster than the fastest YOLO model

Sources:

  • "An Attention-based Convolutional Neural Network With Spatial Transformer Module for Automated Optical Inspection of Small Objects" (IEEE Transactions on Instrumentation and Measurement, 2025;74)
  • Journal of Engineering (May 19, 2025, p 2618)
  • IEEE Transactions on Instrumentation and Measurement (www.ieee.org/)