结合Cannikin’s Law的离线数据增广方法研究

doi:10.3778/j.issn.1002-8331.2106-0241

摘要/Abstract

摘要： 数据增广是提升深度学习模型性能的有效方法之一。针对多类别目标检测任务中检测性能不平衡问题，提出一种针对“短板类别”（检测性能远低于模型平均检测性能的类别）的离线数据增广方法。受Cannikin’s Law的启发，采用基于复制粘贴（copy-paste）机制的场景多样性增广方法。随机采集训练集中“短板类别”实例区域，通过相似性度量机制选取训练集中增广目标样本进行随机粘贴。为了降低随机粘贴导致的遮挡问题，采用基于自遮挡（cut-replace）机制的增广方法提升模型遮挡表达能力。通过截取样本自身区域，对特征表达最显著区域进行遮挡。实验表明，FCOS目标检测框架在PASCAL VOC数据上的平均检测精度（mean average precision，mAP）从79.10%提升到83.90%，其中短板类别更为显著，提升了20.8个百分点。在MS-COCO数据上平均检测精度提升了0.9个百分点。

关键词: 数据增广, Cannikin’s Law, 相似性度量机制, 自遮挡, 目标检测

Abstract: Data augmentation is one of the effective methods to improve the performance of a deep learning model. Aiming at the imbalance of detection performance in multi-class object detection tasks, an offline data augmentation method for “shortboard categories”（categories whose detection performance is far lower than the average detection performance of the model） is proposed. Inspired by Cannikin’s Law, a scene diversity augmentation method based on copy-paste mechanism is adopted. The instance area of the “shortboard category” in the training dataset is randomly collected, and the augmented target samples in the training set are selected by a similarity measurement mechanism for random pasting. In order to reduce the occlusion problem caused by random pasting, an augmented method based on the cut-replace mechanism is used to improve the occlusion representation ability of the model. By intercepting the sample’s own region, the most significant region of feature expression is self-occluded. Experimental results show that the mean average precision（mAP） of FCOS object detection framework on PASCAL VOC data is improved from 79.10% to 83.90%, and the shortboard category is more significant, which is improved by 20.8?percentage points. In MS COCO data, the average detection accuracy is improved by 0.9?percentage points.

Key words: data augmentation, Cannikin’s Law, similarity measurement mechanism, self-occlusion, object detection

邓雪, 赵皓, 张静, 梅菠萍, 张华. 结合Cannikin’s Law的离线数据增广方法研究[J]. 计算机工程与应用, 2023, 59(1): 207-212.

DENG Xue, ZHAO Hao, ZHANG Jing, MEI Boping, ZHANG Hua. Research on Offline Data Augmentation Method Jointed with Cannikin’s Law[J]. Computer Engineering and Applications, 2023, 59(1): 207-212.

参考文献

[1] 许德刚，王露，李凡.深度学习的典型目标检测算法研究综述[J].计算机工程与应用，2021，57（8）：10-25.
XU D G，WANG L，LI F.Review of typical object detection algorithms for deep learning[J].Computer Engineering and Applications，2021，57（8）：10-25.
[2] TYNER S，BRIATTE F，HOFMANN H.Network visualization with ggplot2[J].The R Journal，2017，9（1）：27-59.
[3] SZEGEDY C，IOFFE S，VANHOUCKE V，et al.Inception-v4，inception-ResNet and the impact of residual connections on learning[C]//Proceedings of the AAAI Conference on Artificial Intelligence，2017.
[4] JEGADEESH N，TITMAN S.Momentum[J].Annual Review of Financial Economics，2011，3：493-509.
[5] MARGALIOT M.Root-mean-square gains of switched linear systems：A variational approach[J].Automatica，2008，44（9）：2398-2402.
[6] KISANTAL M，WOJNA Z，MURAWSKI J，et al.Augmentation for small object detection[C]//Proceedings of the 9th International Conference on Advances in Computing and Information Technology（ACITY 2019），2019.
[7] GHIASI G，CUI Y，SRINIVAS A，et al.Simple copy-paste is a strong data augmentation method for instance segmentation[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition，2021：2917-2927.
[8] XIAOZHI C，KAUSTAY KUN D，ZHU Y.3D object proposals for accurate object class detection[C]//Proceedings of the International Conference on Neural Information Processing Systems，2015.
[9] NAJIBI M，SAMANGOUEI P，CHELLAPPA R，et al.SSH：Single stage headless face detector[C]//Proceedings of the IEEE International Conference on Computer Vision，2017：4885-4894.
[10] CAO G，XIE X，YANG W，et al.Feature-fused SSD：Fast detection for small objects[J].arXiv：1709.05054，2017.
[11] MENIKDIWELA M，NGUYEN C，LI H D，et al.CNN-based small object detection and visualization with feature activation mapping[C]//Proceedings of the International Conference on Image and Vision Computing，2017：1-5.
[12] 谌贵辉，陈伍，李忠兵，等.残差卷积注意网络的图像超分辨率重建[J].计算机工程与应用，2021，57（12）：193-200.
CHEN G H，CHEN W，LI Z B，et al.Image super resolution reconstruction based on residual convolution attention network[J].Computer Engineering and Applications，2021，57（12）：193-200.
[13] ZHONG Z，ZHENG L，KANG G L，et al.Random erasing data augmentation[C]//Proceedings of the AAAI Conference on Artificial Intelligence，2020：13001-13008.
[14] DEVRIES T，TAYLOR G W.Improved regularization of convolutional neural networks with cutout[J].arXiv：1708. 04552，2017.
[15] YUN S，HAN D，CHUN S，et al.CutMix：Regularization strategy to train strong classifiers with localizable features[C]//Proceedings of the IEEE/CVF International Conference on Computer Vision，2019：6022-6031.
[16] LIN T Y，GOYAL P，GIRSHICK R，et al.Focal loss for dense object detection[C]//Proceedings of the IEEE International Conference on Computer Vision，2017：2999-3007.
[17] TIAN Z，SHEN C H，CHEN H，et al.FCOS：Fully convolutional one-stage object detection[C]//Proceedings of the IEEE/CVF International Conference on Computer Vision，2019：9626-9635.