Binary Neural Network for Automated Visual Surface Defect Detection.

Area of Science:

• Materials Science
• Computer Vision
• Machine Learning

Background:

• Machine defects can cause catastrophic failures, necessitating rigorous pre-mounting quality assessment.
• Current defect detection networks exhibit high computational complexity, hindering practical manufacturing deployment.
• Binary neural networks offer a solution by constraining weights and activations, reducing computational load.

Purpose of the Study:

• To introduce binary neural networks for surface defect detection.
• To develop efficient binary network architectures (Bi-ShuffleNet, U-BiNet) for manufacturing quality control.
• To evaluate the performance and computational efficiency of these binary networks.

Main Methods:

• Implementation of binary convolution layers and low bitwidth activations.
• Development of Bi-ShuffleNet for defect classification and U-BiNet for defect segmentation.
• Extensive experimentation on the NEU and Magnetic Tile datasets.

Main Results:

• Bi-ShuffleNet achieved comparable defect classification accuracy to traditional networks with at least a 2x reduction in inference complexity.
• U-BiNet demonstrated similar efficiency gains for defect segmentation tasks.
• The proposed binary networks required significantly fewer operations (OPs) with minimal accuracy loss.

Conclusions:

• Binary neural networks are effective for surface defect detection, offering substantial computational savings.
• The Bi-ShuffleNet and U-BiNet models provide a viable solution for real-time manufacturing quality control.
• The study outlines network design principles for defect detection and binary networks.