An efficient multi-task forest fire and smoke detection model

doi:10.1016/j.engappai.2025.111958

An efficient multi-task forest fire and smoke detection model

2025-08-19

0

PRE

AI

DOI：10.1016/j.engappai.2025.111958

Original

Original request for help

Save

Abstract

En 中文

• To begin with, we integrated the Kolmogorov-Arnold Network into YOLOv8's backbone feature extractor and optimized its activation function, which has boosted both model efficiency and the capability of extracting features for complex nonlinear relationships. • Next, a median-enhanced spatial-channel attention module (MECS) was added to the neck. By combining median, global, and max pooling, it refines feature extraction and enhances performance in complex scenes. • Furthermore, dedicated segmentation and detection heads were designed. The pixel-level category information from segmentation aids detection, leading to a marked improvement in accuracy. • Finally, the Normalized Wasserstein Distance (NWD) loss was introduced. This helps better handle small, sparse targets and addresses the challenge of detecting them.

Keywords:

Kolmogorov-Arnold Network

MECS attention module

segmentation-detection heads

Normalized Wasserstein Distance loss

feature extraction

An efficient multi-task forest fire and smoke detection model

Abstract

Journal

Organization