Abstract: Currently, many smoke recognition algorithms suffer from high false alarm rates, partly due to the fact that most existing convolutional neural networks (CNNs) mainly focus on local information in smoke images during feature extraction, neglecting the global features of smoke images. This bias towards local information processing can easily lead to misjudgments when dealing with variable and complex smoke images. To address this issue, it is necessary to capture the global features of smoke images more accurately, thereby improving the accuracy of smoke recognition algorithms. Therefore, this paper propose a dual-branch smoke recognition method, TCF-Net, which combines the Inception and Transformer structures. This model is improved to enrich feature diversity while reducing channel redundancy. Additionally, the self-attention mechanism from Transformer is introduced, combining its ability to learn global context information with CNNs capacity to learn local relative position information. During feature extraction, a feature coupling unit (FCU) is embedded to continuously interact the local features and global information in both branches, maximizing the retention of both local and global information and enhancing the performance of the algorithm. The proposed algorithm can classify video frames into three states: black smoke, white smoke, and no smoke. Experimental results show that the improved network can better extract smoke features, reducing the false alarm rate while increasing the accuracy to 97.8%, confirming the excellent performance of the algorithm.

Key words: smoke recognition, convolutional neural network, self-attention mechanism, feature fusion