基于改进YOLOv5s的火灾烟火检测模型

Fire and Smoke Detection Model Based on Improved YOLOv5s

摘要: 针对目前火灾烟火识别检测模型存在特征提取能力不足和检测精度不高的问题，提出了一种基于改进YOLOv5s的火灾烟火检测模型。通过融合LVC (Learnable Visual Center)模块和LRes (Light Resnet)模块来构建可学习中心残差模块，在保留输入特征信息的同时，提高对烟火边缘特征的学习能力；在YOLOv5s模型C3模块中的单个残差块内构建分层残差连接，用多尺度Res2net模块替换Bottleneck模块，增强全局特征提取能力。对大量数据集进行处理，发现相较于原始YOLOv5s模型，改进后的YOLOv5s模型对火的平均检测精度值提升了3%，全目标平均检测精度值提高了2.2%。

Abstract: In response to the current fire and smoke detection models that suffer from insufficient feature extraction capabilities and low detection accuracy, an improved YOLOv5s fire and smoke detection model is proposed. By integrating the Learnable Visual Center module and the Light Resnet module, a learnable center residual module is constructed, which enhances the learning ability of fire and smoke edge features while retaining the input feature information. A hierarchical residual connection is built within the single residual block of the YOLOv5s model’s C3 module, and the Bottleneck module is replaced with a multi-scale Res2net module to enhance the global feature extraction capability. After processing a large number of datasets, it is found that compared with the original YOLOv5s model, the improved YOLOv5s model has increased the average detection accuracy for fire by 3%, and the average detection accuracy for all targets has been improved by 2.2%.