Abstract: A fire identification algorithm based on the YOLOv5 algorithm is proposed to address the complex interference factors at fire scenes and the relative lag of traditional fire alarm methods. Firstly, the activation function is improved to enhance the model's nonlinear expression and generalization ability, and the loss function is improved to reduce the learning of unnecessary features by the model. Then, the attention mechanism (CBAM) is introduced to enhance the model's bidirectional perception in both channels and space. Finally, a comparative experiment between the improved YOLOv5 algorithm and other recognition algorithms is designed. Qualitative and quantitative analysis proves the effectiveness of the improved algorithm. Compared with the original YOLOv5 algorithm, the model's average precision (AP) has been improved by 5.85%, and the overall network performance has been significantly improved, meeting the requirements for accuracy and real-time detection of fire images. The test results prove its good performance and application value.

Key words: font-family:-apple-system, blinkmacsystemfont, ", font-size:14px, background-color:#FFFFFF, ">computer vision, petrochemical enterprise, fire intelligent monitoring, network structure, activation function, loss function, evaluation index, ablation experiment

摘要： 针对火灾现场干扰因素复杂、传统火灾报警方式相对滞后的问题，提出了一种基于YOLOv5改进的火灾识别算法。首先，通过改进激活函数提升模型非线性表达及泛化能力，改进损失函数，减少模型对非必要特征的学习；然后，通过引入注意力机制CBAM，增强模型在通道和空间的双向感知力；最后，设计改进YOLOv5算法与其他识别算法对比实验。通过定性及定量分析证明改进算法的有效性，与原YOLOv5算法相比，模型平均精度AP提高了5.85%，整体网络性能有了较明显的提升，满足了火灾图像检测准确度、实时性的需求，测试结果证明其具有良好的性能以及应用价值。

关键词: font-family:-apple-system, blinkmacsystemfont, ", font-size:14px, background-color:#FFFFFF, ">计算机视觉, 石化企业, 火灾智能监测, 网络结构, 激活函数, 损失函数, 评价指标, 消融实验