Search

• A new method for multi-scale smoke detection combining CNN and transformer is proposed. • A multiple receptive Fields(MRF) module is designed to improve the ability to extract features from smoke at different scales. • A enhanced pool Former(EPF) module is designed to improve the expression ability of smoke features. • A dataset of multi-scale smoke detection was constructed. Smoke identification in photos has long been difficult because of the intricate variations in smoke's texture, color, and shape. This article suggests a Transformer network and YOLOv5-based smoke detector to solve this problem. In particular, we suggest using the Enhanced Pool Former (EPF) structure to improve the expression ability of smoke features by obtaining greater global information. Additionally, to improve the recognition capacity of slight smoke, we strive to prevent the loss of information by increasing the number of detecting heads and introducing more feature fusion. We do this by using the NWDLoss to compensate for the sensitivity of IoU to the position difference of small objects. In addition, we develop the Multiple Receptive Fields (MRF) module to improve the ability to extract features from smoke at different scales. Our solution outperforms current methods on our custom dataset for AP and AP50, making it well-suited for smoke detection. [ABSTRACT FROM AUTHOR]