《计算机应用》唯一官方网站 ›› 2025, Vol. 45 ›› Issue (7): 2361-2368.DOI: 10.11772/j.issn.1001-9081.2024070959
收稿日期:
2024-07-09
修回日期:
2024-09-29
接受日期:
2024-10-09
发布日期:
2025-07-10
出版日期:
2025-07-10
通讯作者:
宿景芳
作者简介:
王震洲(1978—),男,河北石家庄人,教授,博士,主要研究方向:图像处理、模式识别基金资助:
Zhenzhou WANG1, Fangfang GUO1, Jingfang SU1(), He SU2, Jianchao WANG1
Received:
2024-07-09
Revised:
2024-09-29
Accepted:
2024-10-09
Online:
2025-07-10
Published:
2025-07-10
Contact:
Jingfang SU
About author:
WANG Zhenzhou, born in 1978, Ph. D., professor. His research interests include image processing, pattern recognition.Supported by:
摘要:
输电线路的智能巡检视觉任务对电力系统的安全稳定至关重要。尽管深度学习网络在分布一致的训练和测试数据集上表现良好,但实际应用中数据分布的偏差常常会降低模型性能。为了解决这一问题,提出一种基于对比学习的训练方法(TMCL),旨在增强模型鲁棒性。首先,构建专为输电线路场景设计的基准测试集TLD-C (Transmission Line Dataset-Corruption)用于评估模型在面对图像损坏时的鲁棒性;其次,通过构建对类别特征敏感的正负样本对,提升模型对不同类别特征的区分能力;然后,使用结合对比损失和交叉熵损失的联合优化策略对特征提取过程施加额外约束,以优化特征向量的表征;最后,引入非局部特征去噪网络(NFD)用于提取与类别密切相关的特征。实验结果表明,模型改进后的训练方法在输电线路数据集(TLD)上的平均精度比原始方法高出3.40个百分点,在TLD-C数据集上的相对损坏精度(rCP)比原始方法高出4.69个百分点。
中图分类号:
王震洲, 郭方方, 宿景芳, 苏鹤, 王建超. 面向智能巡检的视觉模型鲁棒性优化方法[J]. 计算机应用, 2025, 45(7): 2361-2368.
Zhenzhou WANG, Fangfang GUO, Jingfang SU, He SU, Jianchao WANG. Robustness optimization method of visual model for intelligent inspection[J]. Journal of Computer Applications, 2025, 45(7): 2361-2368.
类别 | 图像总数 | 检测目标数 |
---|---|---|
绝缘子(insulator) | 370 | 625 |
杆塔(pole tower) | 366 | 549 |
挖掘机(excavator) | 355 | 400 |
塔吊(tower crane) | 356 | 534 |
吊车(crane) | 355 | 375 |
油罐车(tanker truck) | 354 | 388 |
鸟巢(nest) | 352 | 400 |
铲车(forklift) | 352 | 405 |
大卡车(big truck) | 354 | 378 |
推土车(bulldozer) | 352 | 380 |
表1 输电线路数据集的类别和数量
Tab. 1 Categories and numbers in transmission line dataset
类别 | 图像总数 | 检测目标数 |
---|---|---|
绝缘子(insulator) | 370 | 625 |
杆塔(pole tower) | 366 | 549 |
挖掘机(excavator) | 355 | 400 |
塔吊(tower crane) | 356 | 534 |
吊车(crane) | 355 | 375 |
油罐车(tanker truck) | 354 | 388 |
鸟巢(nest) | 352 | 400 |
铲车(forklift) | 352 | 405 |
大卡车(big truck) | 354 | 378 |
推土车(bulldozer) | 352 | 380 |
网络模型 | TLD | TLD-C | |
---|---|---|---|
AP | mCP | rCP | |
AllConvNet | 83.80 | 52.70 | 62.89 |
DenseNet | 81.20 | 51.80 | 63.80 |
WideResNet | 80.70 | 53.10 | 65.80 |
ResNeXt | 82.40 | 54.30 | 65.90 |
ResNet | 81.40 | 53.20 | 65.36 |
表2 不同网络模型的鲁棒性性能 ( %)
Tab. 2 Robustness performance of different network models
网络模型 | TLD | TLD-C | |
---|---|---|---|
AP | mCP | rCP | |
AllConvNet | 83.80 | 52.70 | 62.89 |
DenseNet | 81.20 | 51.80 | 63.80 |
WideResNet | 80.70 | 53.10 | 65.80 |
ResNeXt | 82.40 | 54.30 | 65.90 |
ResNet | 81.40 | 53.20 | 65.36 |
数据增强方法 | AllConvNet | DenseNet | WideResNet | ResNeXt | ResNet |
---|---|---|---|---|---|
Standard | 52.70 | 51.80 | 53.10 | 54.30 | 53.20 |
Cutout | 53.60 | 50.40 | 54.30 | 54.70 | 53.60 |
mixup | 57.50 | 57.40 | 58.50 | 61.20 | 59.70 |
Auto Augment | 54.30 | 55.10 | 57.40 | 58.70 | 57.10 |
AugMix | 62.50 | 61.60 | 63.30 | 64.20 | 63.50 |
本文方法 | 63.10 | 61.90 | 63.50 | 65.20 | 63.80 |
表3 不同数据增强方法在TLD-C数据集上的mCP结果 ( %)
Tab. 3 mCP results of different data augmentation methods on TLD-C dataset
数据增强方法 | AllConvNet | DenseNet | WideResNet | ResNeXt | ResNet |
---|---|---|---|---|---|
Standard | 52.70 | 51.80 | 53.10 | 54.30 | 53.20 |
Cutout | 53.60 | 50.40 | 54.30 | 54.70 | 53.60 |
mixup | 57.50 | 57.40 | 58.50 | 61.20 | 59.70 |
Auto Augment | 54.30 | 55.10 | 57.40 | 58.70 | 57.10 |
AugMix | 62.50 | 61.60 | 63.30 | 64.20 | 63.50 |
本文方法 | 63.10 | 61.90 | 63.50 | 65.20 | 63.80 |
训练方法 | TLD | TLD-C | |
---|---|---|---|
AP | mCP | rCP | |
AugMix | 82.10 | 48.80 | 59.44 |
本文方法 | 83.40 | 52.20 | 62.59 |
表4 不同模型在不同训练方法下的鲁棒性性能 ( %)
Tab. 4 Robustness performance of different models under different training methods
训练方法 | TLD | TLD-C | |
---|---|---|---|
AP | mCP | rCP | |
AugMix | 82.10 | 48.80 | 59.44 |
本文方法 | 83.40 | 52.20 | 62.59 |
模型 | TLD | TLD-C | |
---|---|---|---|
AP | mCP | rCP | |
ResNet-50 | 81.40 | 57.00 | 70.02 |
ResNet-50NFD | 82.60 | 58.60 | 70.94 |
表5 NFD对模型鲁棒性的影响 ( %)
Tab. 5 Influence of NFD on model robustness
模型 | TLD | TLD-C | |
---|---|---|---|
AP | mCP | rCP | |
ResNet-50 | 81.40 | 57.00 | 70.02 |
ResNet-50NFD | 82.60 | 58.60 | 70.94 |
方法 | LS | NFD | TLD | TLD-C | |
---|---|---|---|---|---|
AP | mCP | rCP | |||
MD | — | √ | 82.60 | 55.70 | 67.43 |
ML | √ | — | 83.50 | 57.90 | 69.34 |
MO | — | — | 81.40 | 53.20 | 65.36 |
本文方法 | √ | √ | 84.80 | 59.40 | 70.05 |
表6 消融实验结果 ( %)
Tab. 6 Results of ablation experiments
方法 | LS | NFD | TLD | TLD-C | |
---|---|---|---|---|---|
AP | mCP | rCP | |||
MD | — | √ | 82.60 | 55.70 | 67.43 |
ML | √ | — | 83.50 | 57.90 | 69.34 |
MO | — | — | 81.40 | 53.20 | 65.36 |
本文方法 | √ | √ | 84.80 | 59.40 | 70.05 |
[1] | 卢志博,徐澄宇,杨罡,等.基于改进YOLOv3的输电线路部件实时检测[J].电测与仪表,2023, 60(7): 138-144. |
LU Z B, XU C Y, YANG G, et al. Real-time detection of transmission line components based on improved YOLOv3 [J]. Electrical Measurement and Instrumentation, 2023, 60(7): 138-144. | |
[2] | 刘黎,韩睿,韩译锋,等.改进的Faster-RCNN目标检测方法在变电站悬挂异物检测中的应用[J].电测与仪表,2021, 58(1): 142-146. |
LIU L, HAN R, HAN Y F, et al. Application of an improved Faster-RCNN object detection method in the detection of suspended foreign matters in substation [J]. Electrical Measurement and Instrumentation, 2021, 58(1): 142-146. | |
[3] | HUANG G, LIN Z, VAN DER MAATEN L, et al. Densely connected convolutional networks [C]// Proceedings of the 2017 IEEE Conference on Computer Vision and Pattern Recognition. Piscataway: IEEE, 2017: 2261-2269. |
[4] | TAORI R, DAVE A, SHANKAR V, et al. Measuring robustness to natural distribution shifts in image classification [C]// Proceedings of the 34th International Conference on Neural Information Processing Systems. Red Hook: Curran Associates Inc., 2020: 18583-18599. |
[5] | RECHT B, ROELOFS R, SCHMIDT L, et al. Do ImageNet classifiers generalize to ImageNet? [C]// Proceedings of the 36th International Conference on Machine Learning. New York: JMLR.org, 2019: 5389-5400. |
[6] | DODGE S, KARAM L. A study and comparison of human and deep learning recognition performance under visual distortions [C]// Proceedings of the 26th International Conference on Computer Communication and Networks. Piscataway: IEEE, 2017: 1-7. |
[7] | SHEN H, HU B C, CZARNECKI K, et al. Assessing visually-continuous corruption robustness of neural networks relative to human performance [C]// Proceedings of the 2025 IEEE/CVF Winter Conference on Applications of Computer Vision. Piscataway: IEEE, 2025: 6300-6310. |
[8] | YANG K, YAU J H, LI F F, et al. A study of face obfuscation in ImageNet [C]// Proceedings of the 39th International Conference on Machine Learning. New York: JMLR.org, 2022: 25313-25330. |
[9] | HENDRYCKS D, DIETTERICH T. Benchmarking neural network robustness to common corruptions and perturbations [EB/OL]. [2024-05-24]. . |
[10] | HENDRYCKS D, MU N, CUBUK E D, et al. AugMix: a simple method to improve robustness and uncertainty under data shift [EB/OL]. [2024-05-24]. . |
[11] | XING W, YAO J, LIU Z, et al. Contrastive JS: a novel scheme for enhancing the accuracy and robustness of deep models [J]. IEEE Transactions on Multimedia, 2023, 25: 7881-7893. |
[12] | SAIKIA T, SCHMID C, BROX T. Improving robustness against common corruptions with frequency biased models [C]// Proceedings of the 2021 IEEE/CVF International Conference on Computer Vision. Piscataway: IEEE, 2021: 10191-10200. |
[13] | WANG S, VELDUIS R, BRUNE C, et al. A survey on the robustness of computer vision models against common corruptions [EB/OL]. [2024-08-24]. . |
[14] | JAGANATHAN A P. Meta-styled CNNs: boosting robustness through adaptive learning and style transfer [J/OL]. International Journal of Information Technology, 2024: 1-14 [2024-08-20] . |
[15] | 王彦海,郭宸昕,吴德强.基于改进YOLOv7的输电线路机械外破隐患目标检测方法[J/OL].电测与仪表[2024-09-04]. . |
WANG Y H, GUO C X, WU D Q. Hidden target detection method for mechanical external damage of transmission line based on improved YOLOv7 [J/OL]. Electrical Measurement and Instrumentation [2024-09-04]. . | |
[16] | 龙乐云,周腊吾,刘淑琴,等.改进YOLOv5算法下的输电线路外破隐患目标检测研究[J].电子测量与仪器学报,2022, 36(11): 245-253. |
LONG L Y, ZHOU L W, LIU S Q, et al. Identification of hidden damage targets by external forces based on domain adaptation and attention mechanism [J]. Journal of Electronic Measurement and Instrumentation, 2022, 36(11): 245-253. | |
[17] | 叶翔,孙嘉兴,甘永叶,等.改进YOLOv3模型在无人机巡检输电线路部件缺陷检测中的应用研究[J].电测与仪表,2023, 60(5): 85-91. |
YE X, SUN J X, GAN Y Y, et al. Application of improved YOLOv3 model in defect detection of transmission line components in UAV patrol inspection [J]. Electrical Measurement and Instrumentation, 2023, 60(5): 85-91. | |
[18] | 张重生,陈杰,李岐龙,等.深度对比学习综述[J].自动化学报,2023, 49(1): 15-39. |
ZHANG C S, CHEN J, LI Q L, et al. Deep contrastive learning: a survey [J]. Acta Automatica Sinica, 2023, 49(1): 15-39. | |
[19] | PATACCHIOLA M, STORKEY A. Self-supervised relational reasoning for representation learning [C]// Proceedings of the 34th International Conference on Neural Information Processing Systems. Red Hook: Curran Associates Inc., 2020: 4003-4014. |
[20] | 钱来,赵卫伟.基于对比学习和注意力机制的文本分类方法[J].计算机工程,2024, 50(7): 104-111. |
QIAN L, ZHAO W W. Text classification method based on contrastive learning and attention mechanism [J]. Computer Engineering, 2024, 50(7): 104-111. | |
[21] | 刘传洋,吴一全.基于深度学习的输电线路视觉检测方法研究进展[J].中国电机工程学报, 2023, 43(19): 7423-7446. |
LIU C Y, WU Y Q. Research progress of vision detection methods based on deep learning for transmission lines [J]. Proceedings of the CSEE, 2023, 43(19): 7423-7446. | |
[22] | SPRINGENBERG J T, DOSOVITSKIY A, BROX T, et al. Striving for simplicity: the all convolutional net [EB/OL]. [2024-08-25]. . |
[23] | ZAGORUYKO S, KOMODAKIS N. Wide residual networks [C]// Proceedings of the 2016 British Machine Vision Conference. Durham: BMVA Press, 2016: No.87. |
[24] | XIE S, GIRSHICK R, DOLLÁR P, et al. Aggregated residual transformations for deep neural networks [C]// Proceedings of the 2017 IEEE Conference on Computer Vision and Pattern Recognition. Piscataway: IEEE, 2017: 5987-5995. |
[25] | HE K, ZHANG X, REN S, et al. Deep residual learning for image recognition [C]// Proceedings of the 2016 IEEE Conference on Computer Vision and Pattern Recognition. Piscataway: IEEE, 2016: 770-778. |
[26] | TERMRITTHIKUN C, JAMTSHO Y, MUNEESAWANG P. An improved residual network model for image recognition using a combination of snapshot ensembles and the cutout technique [J]. Multimedia Tools and Applications, 2020, 79(1/2): 1475-1495. |
[27] | ZHANG H, CISSE M, DAUPHIN Y N, et al. mixup: beyond empirical risk minimization [EB/OL]. [2023-10-11]. . |
[28] | CUBUK E D, ZOPH B, MANÉ D, et al. AutoAugment: learning augmentation policies from data [C]// Proceedings of the 2019 IEEE/CVF Conference on Computer Vision and Pattern Recognition. Piscataway: IEEE, 2019: 113-123. |
[1] | 索晋贤, 张丽萍, 闫盛, 王东奇, 张雅雯. 可解释的深度知识追踪方法综述[J]. 《计算机应用》唯一官方网站, 2025, 45(7): 2043-2055. |
[2] | 谢劲, 褚苏荣, 强彦, 赵涓涓, 张华, 高勇. 用于胸片中硬负样本识别的双支分布一致性对比学习模型[J]. 《计算机应用》唯一官方网站, 2025, 45(7): 2369-2377. |
[3] | 齐巧玲, 王啸啸, 张茜茜, 汪鹏, 董永峰. 基于元学习的标签噪声自适应学习算法[J]. 《计算机应用》唯一官方网站, 2025, 45(7): 2113-2122. |
[4] | 赵小阳, 许新征, 李仲年. 物联网应用中的可解释人工智能研究综述[J]. 《计算机应用》唯一官方网站, 2025, 45(7): 2169-2179. |
[5] | 姜超英, 李倩, 刘宁, 刘磊, 崔立真. 基于图对比学习的再入院预测模型[J]. 《计算机应用》唯一官方网站, 2025, 45(6): 1784-1792. |
[6] | 余明峰, 秦永彬, 黄瑞章, 陈艳平, 林川. 基于对比学习增强双注意力机制的多标签文本分类方法[J]. 《计算机应用》唯一官方网站, 2025, 45(6): 1732-1740. |
[7] | 颜文婧, 王瑞东, 左敏, 张青川. 基于风味嵌入异构图层次学习的食谱推荐模型[J]. 《计算机应用》唯一官方网站, 2025, 45(6): 1869-1878. |
[8] | 花天辰, 马晓宁, 智慧. 基于浅层人工神经网络的可移植执行恶意软件静态检测模型[J]. 《计算机应用》唯一官方网站, 2025, 45(6): 1911-1921. |
[9] | 李岚皓, 严皓钧, 周号益, 孙庆赟, 李建欣. 基于神经网络的多尺度信息融合时间序列长期预测模型[J]. 《计算机应用》唯一官方网站, 2025, 45(6): 1776-1783. |
[10] | 龙雨菲, 牟宇辰, 刘晔. 基于张量化图卷积网络和对比学习的多源数据表示学习模型[J]. 《计算机应用》唯一官方网站, 2025, 45(5): 1372-1378. |
[11] | 牛四杰, 刘昱良. 基于知识蒸馏双分支结构的视网膜病变辅助诊断方法[J]. 《计算机应用》唯一官方网站, 2025, 45(5): 1410-1414. |
[12] | 陈凯, 叶海良, 曹飞龙. 基于局部-全局交互与结构Transformer的点云分类算法[J]. 《计算机应用》唯一官方网站, 2025, 45(5): 1671-1676. |
[13] | 胡文彬, 蔡天翔, 韩天乐, 仲兆满, 马常霞. 融合对比学习与情感分析的多模态反讽检测模型[J]. 《计算机应用》唯一官方网站, 2025, 45(5): 1432-1438. |
[14] | 王丹, 张文豪, 彭丽娟. 基于深度学习的智能反射面辅助通信系统信道估计[J]. 《计算机应用》唯一官方网站, 2025, 45(5): 1613-1618. |
[15] | 王文鹏, 秦寅畅, 师文轩. 工业缺陷检测无监督深度学习方法综述[J]. 《计算机应用》唯一官方网站, 2025, 45(5): 1658-1670. |
阅读次数 | ||||||
全文 |
|
|||||
摘要 |
|
|||||