CFI-YOLO: A Lightweight Algorithm for Fire Detection

In recent years, with the rapid advancement of urbanization, fire incidents have occurred with increasing frequency, posing a serious threat to public safety and property. To address the challenges of insufficient detection accuracy and high model complexity in existing flame detection algorithms, this paper proposes a lightweight flame detection algorithm based on the YOLOv8 framework, named CFI-YOLO. First, this paper proposes a novel FasterEMA module that integrates the advantages of FasterNet and EMA, effectively enhancing the network’s multi-scale feature learning capability and detection accuracy. Second, a Cross-scale Feature Fusion Module (CCFM) is introduced into the neck network, which employs lightweight convolution operations and a unified channel dimension strategy to reduce the model’s parameter count and computational complexity significantly. Third, the loss function is optimized using the Inner-EIoU, which incorporates an internal overlap penalty mechanism and dynamic non-maximum suppression, leading to a marked improvement in the localization accuracy for small-scale flames. Experiments conducted on a self-constructed dataset demonstrate that CFI-YOLO achieves a parameter size of only 1.9M, representing 61.3% of YOLOv8, while attaining a Precision of 90.2%, an improvement of 2.9%. The [email protected] reaches 88.8%, an increase of 1.0%. Comparative evaluations further validate the superior efficiency and accuracy of CFI-YOLO, making it well-suited for flame detection in practical applications.