A Hybrid Network of Transformer and CNN by Attention-Driven Edge Enhancement for Smoke Segmentation

• Dena Zhu,
• Zhe Yang,
• Mingjin Wang,
• Lei Zhang,
• Kang Li,
• Feiniu Yuan,

Vol. 19, No. 11, November 30, 2025
10.3837/tiis.2025.11.013, Download Paper (Free):

• CNN
• Edge enhancement
• Interactive attention
• Smoke segmentation
• Transformer

Abstract

In the early stage of a fire, timely detection of smoke is crucial for firefighters to plan rescue and firefighting operations, which can significantly reduce property damage and casualties. However, due to limited monitoring range and slow response speed, traditional smoke detection is not suitable for open space such as forests. To address this challenge, we propose an efficient Dual-Attention Interactive Transformer-CNN Fusion Network (DAIFNet) for smoke segmentation. DAIFNet successfully integrates Transformer and CNN encoders through unique interactive attention mechanism. An Interactive Spatial Attention Module utilizes asymmetric convolutions to capture more subtle spatial variations, enabling dynamic integration of spatial information between the dual encoders. Concurrently, the Interactive Channel Attention Module allows the dual encoders to focus more on channels with rich information, thereby enhancing feature extraction capabilities. We specifically design an Edge Enhancement Reconstruction Decoder for generating final outputs. The efficient feature decoder is a collaborative method of Edge Enhancement Reconstruction and Decoder Bottleneck Layer. This decoder can restore image edges more accurately, and effectively utilize deep network features for enhancing computational efficiency, reducing information loss and ensuring the prediction of image details. Experimental results fully demonstrate the superior capability of DAIFNet on the SYN70K and smoke fire segmentation (SFS3K) datasets. On these two datasets, DAIFNet achieves 79.3% and 80.2% of mIoUs, respectively. The parameter number is only 9.16M, and the computational complexity is as low as 3.06G FLOPs. This reflects its efficiency and practicality for real-world applications.

Statistics

Cite this article