摘要
为了提高钢铁质量图像检测的效率和精度,提高生产自动化水平,该文提出一种改进的Res-UNet网络分割算法。使用ResNet50代替ResNet18作为编码模块,增强特征提取能力;修改编码模块,使残差块间稠密连接,增强浅层特征的深度延展,充分利用特征;使用加权Dice损失和加权交叉熵损失(BCEloss)结合的新损失函数缓解样本不均衡的情况;数据集增强策略保证网络学习更多的样本特征,增强细节分割精度。相比于经典的UNet算法,组合优化后的Res-UNet网络的Dice系数最多提高了12.64%,达到0.7930,网络训练时间更短,对各类缺陷的分割精准度更优,证明该文算法在钢铁表面缺陷分割领域具有应用价值。
In order to improve the efficiency and accuracy of steel quality images detection and promote the automation level of industry,an improved Res-UNet segmentation algorithm is proposed.ResNet50 is used instead of ResNet18 as the encode module to enhance feature extraction capability.Structure like DenseNet is added to encode module,which helps to make full use of shallow feature maps.A new loss function combining weighted Dice loss and weighted Binary Cross Entropy loss(BCEloss)is used to alleviate data imbalance.Data set enhancement strategy ensures that the network learns more features and improves the segmentation accuracy.Compared with the classic UNet,the Dice coefficient of the improved Res-UNet increases by 12.64%and reaches 0.7930.In all,the improved Res-UNet achieves much better accuracy on various defects while requires much less training efforts.The algorithm proposed by this paper is of practical use in the field of steel surface defect segmentation.
作者
李原
李燕君
刘进超
范衠
王庆林
LI Yuan;LI Yanjun;LIU Jinchao;FAN Zhun;WANG Qinglin(School of Automation,Beijing Institute of Technology,Beijing 100081,China;College of Artificial Intelligence,Nankai University,Tianjin 300071,China;College of Engineering,Shantou University,Shantou 515063,China)
出处
《电子与信息学报》
EI
CSCD
北大核心
2022年第5期1513-1520,共8页
Journal of Electronics & Information Technology
作者简介
通信作者:李原:男,1977年生,副教授,研究方向为智能机器人系统、计算机视觉、人工智能,liyuan@bit.edu.cn;李燕君:女,1997年生,硕士生,研究方向为计算机视觉、深度学习;刘进超:男,1981年生,副教授,研究方向为机器学习、机器视觉、智能检测;范衠:男,1974年生,教授,研究方向为人工智能与机器人、智能计算、图像处理;王庆林:男,1963年生,教授,研究方向为智能信息处理、非线性控制.
