基于UNet模型燃气管道高后果区分割方法研究  

Research on segmentation method for high-consequence areas of gas pipeline based on UNet model

作　　者：凌晓[1] 王昕越 郭凯[1] 孙宝财 程凌宇 LING Xiao;WANG Xinyue;GUO Kai;SUN Baocai;CHENG Lingyu(College of Petrochemical Technology,Lanzhou University of Technology,Lanzhou Gansu 730050,China;Gansu Special Equipment Inspection and Testing Institute,Lanzhou Gansu 730050,China)

机构地区：[1]兰州理工大学石油化工学院,甘肃兰州730050 [2]甘肃省特种设备检验检测研究院,甘肃兰州730050

出　　处：《中国安全生产科学技术》2024年第4期157-162,共6页Journal of Safety Science and Technology

基　　金：国家自然科学基金项目(52204074);甘肃省科技计划项目(23YFGA0059)。

摘　　要：为提升燃气管道设施监测和事故应急响应中的高后果区图像分割精准度和可靠性,通过改进UNet模型结构,使用优化后的Inception Block模块、通道注意力和空间注意力机制的方法,提升模型捕捉关键特征的能力,并引入高斯噪声增强模型鲁棒性,采用保留最佳参数策略得到最优训练参数。然后对SE UNet、UNet++、原始UNet以及改进后UNet模型在航拍图像数据集上的分割效果进行对比和分析。研究结果表明:相对SE UNet、UNet++和原始UNet,改进后UNet模型在分割效果上表现更佳,综合性能优于其他模型。同时,改进后UNet模型提高了分割准确性,降低了误检和漏检风险。研究结果可为燃气管道设施的安全运行和维护提供有力支持。In order to improve the accuracy and reliability of high-consequence area image segmentation in gas pipeline facility monitoring and emergency response,the UNet model was improved and optimized After the InceptionBlock module,channel attention and spatial attention mechanism methods,the model’s ability to capture key features is improved,and Gaussian noise is introduced to enhance the model robustness The optimal training parameters are obtained by using the strategy of preserving the best parameters.Then,the segmentation effects of SE UNet,UNet++,original UNet and improved UNet models on aerial image data sets are compared and analyzed.The results show that compared with SE UNet,UNet++and the original UNet,the improved UNet model is efficient in segmentation The results show better performance,and the comprehensive performance is better than other models.At the same time,the improved UNet model improves the segmentation accuracy and reduces the risk of false detection and missing detection.The results can be flammable Provide strong support for the safe operation and maintenance of gas pipeline facilities.

关 键 词：深度学习 UNet模型 卷积神经网络 高后果区 图像分割 

分 类 号：X937[环境科学与工程—安全科学]