文献类型：期刊文献

中文题名：面向小样本学习的石化管道缺陷分割模型设计和剪枝加速

英文题名：Petrochemical Pipeline Defect Segmentation Model Design and Pruning Acceleration for Few-shot Learning

作者：甘洲洋[1];任温琦[1];赵超强[1];唐漾[1];钱锋[1]

机构：[1]华东理工大学信息科学与工程学院,上海200237

年份：2026

卷号：33

期号：3

起止页码：520

中文期刊名：控制工程

外文期刊名：Control Engineering of China

收录：;北大核心:【北大核心2023】；

基金：国家科技部重点研发计划项目(2021YFB1714300);中国石油科技创新基金研究项目(2021D002-0902)。

语种：中文

中文关键词：石化管道;小样本学习;缺陷分割;网络剪枝

外文关键词：Petrochemical pipeline;few-shot learning;defect segmentation;network pruning

摘要：石化管道的缺陷分割对管道的稳定运行至关重要,但石化管道的缺陷分割技术仍然存在训练样本稀缺、实时性差的问题。为此,提出一种实时小样本分割方法。首先,提出一种改进的小样本分割模型——基于背景的自适应超像素匹配网络(adaptive superpixelguided network with background, ASGNet-BG),通过引入管道的背景信息辅助分割,提高了缺陷分割的精度;同时,考虑到模型的实时推理能力,设计适用于小样本分割的双阶段剪枝方法,实现了模型的压缩和推理加速。在石化管道缺陷数据集上的实验结果表明,改进后的ASGNet-BG的平均分割精度达到了64%,优于其他小样本分割模型。剪枝后的模型的内存消耗降低了28%,浮点计算量降低了50%,推理速度提高为原来的3倍,整体分割精度仅损失2.19%,表明所提方法能够有效完成石化管道缺陷的实时分割任务。
Defect segmentation of petrochemical pipeline is very important to maintain the long-term stable operation of petrochemical pipeline.To solve the problems of small amount of defect samples and high real-time detection requirements,a real-time few-shot segmentation method is introduced.Firstly,an improved few-shot segmentation model,ASGNet-BG,is introduced to improve the accuracy of defect segmentation by adding background information.At the same time,a two-stage pruning method is introduced to compress the model and improve the inference speed.The feasibility of the method is verified by experiments on the petrochemical pipeline defect dataset.Experimental results show that the average segmentation accuracy of the improved ASGNet-BG reaches 64%,which is better than other few-shot segmentation models.After pruning,the memory consumption is reduced by 28%,the amount of floating point calculation is reduced by 50%,the inference speed is nearly 3 times faster,and the overall accuracy loss is only 2.19%.The results show that the proposed method can effectively achieve real-time and precise segmentation of petrochemical pipeline defects.