C-ViT: An Improved ViT Model for Multi-Label Classification of Bamboo Chopstick Defects.

Area of Science:

• Computer Vision
• Machine Learning
• Quality Control

Background:

• Disposable bamboo chopstick quality impacts consumer experience and safety.
• Current quality inspection methods require refinement for production demands.
• Vision Transformer (ViT) models face challenges with extreme aspect ratios, like bamboo chopsticks.

Purpose of the Study:

• To develop an improved ViT model (C-ViT) for multi-label defect classification of bamboo chopsticks.
• To introduce a novel Hard Examples Contrastive Loss (HCL) function to enhance learning from difficult examples.
• To improve the accuracy and efficiency of bamboo chopstick quality inspection.

Main Methods:

• Proposed C-ViT model featuring a Convolutional Neural Network Feature Extraction (CFE) module to replace traditional patch embedding.
• Developed Hard Examples Contrastive Loss (HCL) function for dynamic hard example selection and contrastive learning.
• Evaluated models on a self-built bamboo chopstick defect dataset (BCDD) and the public VOC2012 dataset.

Main Results:

• C-ViT improved mean Average Precision (mAP) by 1.2% to 92.8% on the BCDD dataset compared to the ViTS model.
• The addition of HCL further boosted performance to 94.3% mAP on the BCDD dataset.
• The HCL function demonstrated effectiveness in multi-label classification tasks on the VOC2012 dataset.

Conclusions:

• The C-ViT model, with its CFE module, is effective for defect detection in objects with extreme aspect ratios.
• The HCL function significantly enhances multi-label classification performance by focusing on hard examples.
• The proposed methods offer a promising approach for improving automated quality inspection in manufacturing.