Scholarlink

Early and accurate fire detection is important, making fire detection systems essential for reducing damage and ensuring timely response. This work focuses on the use of semantic segmentation for accurate fire detection while meeting the real-time performance requirement. We employ a teacher-student framework, where the teacher model involves MobileNetV2 as a lightweight pretrained encoder, Kolmogorov-Arnold Networks (KAN) for adaptive representation learning, and a Long Short-Term Memory (LSTM) module for temporal awareness, all within a U-Net framework; and the same model, excluding the LSTM module, serves as the student to enable efficient deployment. We also create a new dataset with 1,723 labeled fire images extracted from 57 videos and 605 non-fire images, all collected from diverse scenes and sources. Assessed on the Chino et al.. [2] dataset, the proposed architecture proves its leading performance compared to previous methods, reaching 81.63 in Mean IoU. Additionally, real-time detection also comes into play where it operates at 147.02 frames per second (FPS). This study reveals that through effective knowledge distillation, a lightweight student model achieves superior performance while being dramatically more efficient than its teacher, requiring 94% fewer parameters (0.73M vs 12.61M) and 97% fewer MFLOPs (1349.30 vs 51407.73).

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Distilling Temporal Knowledge into a Spatially Efficient Network for Fire Segmentation: An Approach Involving Kolmogorov-Arnold Networks

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Le Ba Dac,
  • Nguyen Thanh Quynh Tien,
  • Pham Thanh Dat,
  • Nguyen Minh Nhut,
  • Nguyen Dinh Thuan

摘要

Early and accurate fire detection is important, making fire detection systems essential for reducing damage and ensuring timely response. This work focuses on the use of semantic segmentation for accurate fire detection while meeting the real-time performance requirement. We employ a teacher-student framework, where the teacher model involves MobileNetV2 as a lightweight pretrained encoder, Kolmogorov-Arnold Networks (KAN) for adaptive representation learning, and a Long Short-Term Memory (LSTM) module for temporal awareness, all within a U-Net framework; and the same model, excluding the LSTM module, serves as the student to enable efficient deployment. We also create a new dataset with 1,723 labeled fire images extracted from 57 videos and 605 non-fire images, all collected from diverse scenes and sources. Assessed on the Chino et al.. [2] dataset, the proposed architecture proves its leading performance compared to previous methods, reaching 81.63 in Mean IoU. Additionally, real-time detection also comes into play where it operates at 147.02 frames per second (FPS). This study reveals that through effective knowledge distillation, a lightweight student model achieves superior performance while being dramatically more efficient than its teacher, requiring 94% fewer parameters (0.73M vs 12.61M) and 97% fewer MFLOPs (1349.30 vs 51407.73).