Enhanced YOLOv7-Tiny for Small-Scale Fire Detection via Multi-Scale Channel Spatial Attention and Dynamic Upsampling

Fire detection remains a challenging task, particularly in distinguishing small smoke and flames from complex backgrounds. This study proposes an enhanced YOLOv7-tiny model tailored for small-scale fire detection. The model incorporates a Multi-Scale Channel Spatial Attention (CAS-ELAN) module, which combines Convolutional Additive Self-attention (CAS) and Efficient Layer Aggregation Networks (ELAN) to extract discriminative fire features and suppress irrelevant background information. To improve detection performance for small targets, we introduce the DySample module, a dynamic upsampling strategy that enhances feature map resolution. Furthermore, the Focaler-IoU loss function is employed to focus the model on hard samples, enhancing the accuracy of small fire detection. Extensive experiments on a self-collected dataset demonstrate that our method achieves a performance improvement of 3.49% in mAP@0.5 and 6.62% in APs over the baseline YOLOv7-tiny model. These results highlight the potential of our model as a competitive tool for detecting fires, especially at small scales. The source code and datasets are publicly available to facilitate reproducibility and further research in this domain. Code and datasets are available at https://github.com/exoplantlalala/DCF-YOLO