关键词: 矿井外因火灾, YOLOv9m, 目标检测, 轻量化, 注意力机制

Abstract:

To address the problem of strong background interference, slow detection speed, and a high missed-detection rate in the detection of exogenous mine fires, an improved YOLOv9m-based method for object detection and image recognition was proposed. Firstly, the convolution modules in the backbone network of the original model were replaced with the lightweight paddlepaddle lightweight convolutional network (PP-LCNet) module, which reduced the number of model parameters. Secondly, the fully contextual attention (FCAttention) module was embedded into the backbone network to enhance flame feature interaction and optimize weight allocation, thereby improving the accuracy of feature selection and information fusion in complex scenes. Finally, the content-ware reassembly of features (CARAFE) dynamic upsampling operator was introduced into the neck network. Through a content-aware mechanism, feature information was dynamically reorganized, and the accuracy of detail representation was improved. The results show that, compared with the original model, the number of parameters and floating-point operations of the improved model are reduced by 13.2% and 13.6%, respectively, while precision, mean average precision (mAP), and frames per second(FPS) are increased by 5.2%, 3.2%, and 23.2%, respectively. The improved YOLOv9m algorithm ensures real-time fire detection accuracy in complex underground environments. It also significantly improves the detection accuracy in small-target fire-source scenarios. The proposed method meets the requirements of lightweight design and real-time performance, and it provides support for early fire warning and rapid emergency response.

Key words: external mine fire, YOLOv9m, object detection, lightweight, attention mechanism