High Definition Flame Image Recognition Technology based on Improved YOLOv8

Aiming at the problem that conventional YOLOv8 model cannot recognize high definition flame images taken by UAV, this paper proposes a high definition flame recognition method based on the improved YOLOv8 model. By introducing GhostNetV2 module into the YOLOv8 backbone network, this method greatly increases the model parameters and improves the feature extraction capability and flame recognition capability of the model while speeding up the fire detection speed. The Efficient Local Attention (ELA) mechanism is added to further enhance the perception of flame characteristics. By introducing the Slicing Aided Hyper Inference(SAHI) Framework , the detection rate of small targets has been significantly improved. Furthermore, we have enhanced SAHI based on this, resulting in a substantial increase in its speed. The test results show that the performance of the improved YOLOv8 intersects with the YOLOv8n model, and the mAP50 and MAP50-95 improve by 4.0% and 6.5% respectively. Compared with SAHI, the improved SAHI speed increases by 151.97% on average, which effectively realizes the balance between model speed and accuracy, and has broad application value in UAV firefighting tasks.