Multi Scale Forest Fire Small Target Detection Technology

A YOLOv5 based improved model for forest fire target detection is proposed to address the issues of low detection and recognition accuracy caused by multi-scale characteristics, variable morphology, and insufficient discriminative features of small targets in forest fire smoke images. This model changes the pooling structure of the original network to Softpool pooling in the SPP module; Introducing DIoU_NMS replaces NMS and sets the loss function to CIoU+DIoU_NMS, finally optimize the initial anchor box using PSO fusion K-means algorithm and change the number of final output channels. The experimental results show that the average detection accuracy of the improved YOLOv5 for forest fire target detection is 3.4% higher than the original YOLOv5, the above methods can effectively and accurately identify forest fire targets.