BCR-Net: Boundary-Category Refinement Network for Weakly Semi-Supervised X-Ray Prohibited Item Detection with Points


{getToc} $title={Table of Contents}

Summary

BCR-Net is a novel network for Weakly Semi-Supervised X-Ray Prohibited Item Detection with Points (WSSPID-P). It introduces two key modules: Boundary Refinement (BR) module and Category Refinement (CR) module, to address the problems of imprecise localization and inaccurate classification.

Highlights

  • BCR-Net achieves effective feature learning under limited annotations by combining box annotations and point annotations.
  • The BR module performs dual attention to enhance boundary information.
  • The CR module performs scale- and rotation-aware contrastive learning to improve intra-class compactness and inter-class separability.
  • BCR-Net demonstrates significant performance improvements over state-of-the-art WSSPID-P detectors.
  • The model achieves better performance when the CR module or the BR module is used.
  • The contrastive branch in the RPN and RoI heads facilitates the generation of high-quality proposals.
  • The scale- and rotation-aware contrastive loss assists the model in effectively learning object orientation changes.

Key Insights

  • The BR module and the CR module are designed to address the problems of imprecise localization and inaccurate classification in WSSPID-P, and their combination leads to improved performance.
  • The use of dual attention in the BR module enables the model to focus on both the boundaries and salient areas of prohibited items, leading to more accurate localization.
  • The incorporation of contrastive learning in the CR module enhances the model's ability to distinguish between different categories of prohibited items, leading to improved classification accuracy.
  • The scale- and rotation-aware contrastive loss is effective in handling variations in scale and rotation, which is crucial for detecting prohibited items in X-ray images.
  • The contrastive branch in the RPN and RoI heads plays a crucial role in generating high-quality proposals, which is essential for accurate object detection.
  • BCR-Net's ability to achieve effective feature learning under limited annotations makes it a promising approach for WSSPID-P.
  • The experimental results demonstrate the effectiveness of BCR-Net in addressing the problems of imprecise localization and inaccurate classification in WSSPID-P.



Mindmap


Citation

Wong, S. (2024). BCR-Net: Boundary-Category Refinement Network for Weakly Semi-Supervised X-Ray Prohibited Item Detection with Points (Version 1). arXiv. https://doi.org/10.48550/ARXIV.2412.18918

Previous Post Next Post

Contact Form