Shizuoka University
Okabe Lab
日本語

A Study on Improving Fire and Smoke Detection Accuracy
by Introducing an Attention Mechanism Using YOLOv8

Ken Aoki

Shizuoka University

Makoto Okabe

Shizuoka University

研究のTeaser画像

Left: GT (Ground Truth), Center: Inference results of YOLOv8, Right: Inference results of the proposed method. It can be seen that the smoke, which YOLOv8 failed to detect, is successfully detected by the proposed method (YOLOv8 enhanced with the custom ESCFBlock attention mechanism).

Abstract

Video-based fire detection has been attracting significant attention. However, missed detections and false alarms remain major challenges, necessitating further improvements in detection accuracy. In this study, we aim to enhance fire detection accuracy by integrating our original attention mechanism, ESCFBlock, into the network of YOLO, an object detection model known for its excellent real-time performance. Experimental results using the D-fire dataset demonstrated that inserting the ESCFBlock into the P5 layer of the Head is the most effective approach. In particular, the confirmed improvement in Recall represents a practical achievement in minimizing missed fire detections.

Paper

Bachelor Thesis (2026)

IPSJ SIG-CGVI, 201th conference

Video

Material

Bachelor Thesis Presentation

Citation

  • Ken Aoki, Makoto Okabe
    A Study on Improving Fire and Smoke Detection Accuracy by Introducing an Attention Mechanism Using YOLOv8
    Bachelor Thesis, Feb 2026
  • Ken Aoki, Makoto Okabe
    A Study on Improving Fire and Smoke Detection Accuracy by Introducing an Attention Mechanism Using YOLOv8
    IPSJ SIG-CGVI, 201th conference, March 2026