基于自适应多目标进化CNN的图像分割方法  

An image segmentation method based on adaptive multi-objective evolutionary CNN

作　　者：王维[1,2] 王显鹏[2,3] 宋相满 WANG Wei;WANG Xian-peng;SONG Xiang-man(Frontier Science Center for Industrial Intelligence and Systems Optimization,Northeastern University,Shenyang 110819,China;Key Laboratory of Data Analytics and Optimization for Smart Industry,Northeastern University,Shenyang 110819,China;Liaoning Engineering Laboratory of Data Analytics and Optimization for Smart Industry,Shenyang 110819,China)

机构地区：[1]东北大学工业智能与系统优化国家级前沿科学中心,沈阳110819 [2]东北大学智能工业数据解析与优化教育部重点实验室,沈阳110819 [3]辽宁省智能工业数据解析与优化工程实验室,沈阳110819

出　　处：《控制与决策》2024年第4期1185-1193,共9页Control and Decision

基　　金：国家自然科学基金重大项目(72192830,72192831);国家自然科学基金面上项目(62073067);国家111项目(B16009);中央高校基本科研业务费专项资金项目(N2128001)。

摘　　要：卷积神经网络已经成为强大的分割模型,但通常为手动设计,这需要大量时间并且可能导致庞大而复杂的网络.人们对自动设计能够准确分割特定领域图像的高效网络架构越来越感兴趣,然而大部分方法或者没有考虑构建更加灵活的网络架构,或者没有考虑多个目标优化模型.鉴于此,提出一种称为AdaMo-ECNAS的自适应多目标进化卷积神经架构搜索算法,用于特定领域的图像分割,在进化过程中考虑多个性能指标并通过优化模型的多目标适应特定的数据集.AdaMo-ECNAS可以构建灵活多变的预测分割模型,其网络架构和超参数通过基于多目标进化的算法找到,算法基于自适应PBI实现3个目标进化问题,即提升预测分割的F1-score、最大限度减少计算成本以及最大限度挖掘额外训练潜能.将AdaMo-ECNAS在两个真实数据集上进行评估,结果表明所提出方法与其他先进算法相比具有较高的竞争性,甚至是超越的.Convolutional neural networks(CNNs)have become powerful segmentation models,but are usually designed manually,which requires extensive time and can result in large and complex networks.There is a growing enthusiasm for automatically designing efficient architectures that can accurately segment domain-specific images.However,most approaches either do not consider building more flexible network architectures or do not consider multiple objectives to optimize model.To cope with these issues,we propose an adaptive multi-objective evolutionary convolutional neural architecture search algorithm called AdaMo-ECNAS for domain-specific image segmentation,which considers multiple performance metrics during the evolutionary process and adapts to a specific dataset by optimizing multiple objectives.AdaMo-ECNAS can build flexible and versatile predictive segmentation models whose architecture and hyperparameters are found by a multi-objective evolutionary algorithm that is adaptively PBI-based tri-objective evolutionary to improve the F1-score of predictive segmentation,minimize the computational cost and maximize the additional training potential.The main contribution of this work is a fully exploited feature information of segmentation and automatic search for high-performance and efficient architectural models.AdaMo-ECNAS is evaluated on two real world datasets,and the proposed method is competitive and even superior to some advanced algorithms.

关 键 词：卷积神经网络 神经架构搜索 多目标优化问题 基于分解的多目标进化算法 自适应 图像分割 

分 类 号：TP273[自动化与计算机技术—检测技术与自动化装置]