YOLO-MFX: lightweight YOLO with improved flame detection for small targets

A lightweight flame detection model based on YOLOv8n enhancement is presented to address the current issue of many flame detection parameters and low flame recognition accuracy, which is unsuitable for embedded devices. Using the MobileNetV3 model as the backbone network, the model reduces the number of parameters while improving the information about flame features. It then employs a new module called IGSConv, which has a Triplet Attention mechanism that enhances the model’s perception of the flame texture and reduces the flame’s exposure to smoke interference. Afterward, it uses AIFI to enhance the ability of intra- and inter-scale feature interactions to prevent needless feature interactions and to lighten the flame detection model further. The enhanced network model gets 62.8% mAP on the dataset, 3.4% higher than the original model.