基于MAML算法的YOLOv3目标检测模型

沈震宇; 朱昌明; 王喆

doi:10.14135/j.cnki.1006-3080.20201128002

基于MAML算法的YOLOv3目标检测模型

doi: 10.14135/j.cnki.1006-3080.20201128002

沈震宇^1,,
朱昌明^1, ,,
王喆^2, ,

1.
上海海事大学信息工程学院，上海 202010
2.
华东理工大学信息科学与工程学院，上海 200237

基金项目: 晨光计划（18CG54）；中国博士后科学基金（2019M651576）；国家自然科学青年基金资助项目（61602296）；上海市自然科学基金（16ZR1414500）

详细信息

作者简介:
沈震宇（1996—），男，上海嘉定人，硕士生，主要研究方向为目标检测。E-mail： 564669672@qq.com

通讯作者:
朱昌明，E-mail： cmzhu@shmtu.edu.cn

王喆，E-mail： wangzhe@ecust.edu.cn

中图分类号: TP391.4
计量
- 文章访问数: 243
- HTML全文浏览量: 146
- PDF下载量: 10
- 被引次数: 0
出版历程
- 收稿日期: 2020-11-28
- 网络出版日期: 2021-03-24

YOLOv3 Object Detection Model Based on MAML Algorithm

SHEN Zhenyu^1
,,
ZHU Changming^{1
, ,},
WANG Zhe^{2
, ,}

1.
College of Information Engineering, Shanghai Maritime University, Shanghai, China
2.
School of Information Science and Engineering, East China University of Science and Technology, Shanghai 200237, China

摘要

摘要: 作为典型的一体化卷积神经网络，YOLOv3模型的网路传输途径简单，检测速度相对较快，但检测精度较低。当遇到新的目标在训练数据集中存在的样本较少时，模型检测会更加不准确，甚至会出现检测不到的情况。本文根据与模型不相关的元学习算法（Model-Agnostic Meta-Learning，MAML），改进了YOLOv3主干网络的结构，使其具有内循环和外循环的梯度下降，通过在初始参数基础上进行多步的梯度调整，达到仅用小样本数据就能快速收敛的目的。实验结果表明，该方法使得YOLOv3模型的检测精度提升了5.24%，且可以使梯度下降保持稳定，有效地满足YOLOv3模型在小样本数据训练情况下识别目标位置的精准性和泛化性。
- 计算机视觉 /
- 图像识别 /
- 特征提取 /
- 目标检测 /
- 小样本学习
Abstract: Object detection has been a research hotspot in the field of computer vision in recent years. Due to the extensive application of deep learning, the target detection technology combined with deep learning is also developing year by year and making continuous breakthroughs. In the field of target detection, few methods can solve the problem of target detection with few sample categories, and can detect small targets with high accuracy on the basis of training with few sample categories in the detection algorithm of integrated convolutional network. According to the Model-Agnostic Meta-Learning(MAML) algorithm in meta learning, this paper improves the information transmission form of the backbone network in YOLOv3, so that Darknet-53 can achieve two stages of parameter internal update and external update in the aspect of gradient descent. Through its loss function in the case of multi-step gradient adjustment, the weight parameters trained by the model can focus more on the feature information of the target. Even when learning a small number of sample categories and encountering a new task, it can also maintain the sensitivity to the target in the new task. The experiment shows that the mean Average Precision(mAP) value of YOLOv3 model is 74.81%, and the mAP value of YOLOv3 model improved based on MAML algorithm can reach 80.05%, which improves the accuracy by about 5%. The network structure and training mechanism of YOLOv3 improved on the basis of MAML can improve the accuracy of detection in training and prediction, and the weight parameters trained finally can make the model high detection accuracy and high generalization.
- computer vision /
- image recognition /
- feature extraction /
- object detection /
- few shot learning

HTML全文

图 1 MAML训练网络结构流程图

Figure 1. Structure of MAML training network

下载: 全尺寸图片幻灯片

图 2 YOLOv3的样本分类结果

Figure 2. Classification results of YOLOv3

下载: 全尺寸图片幻灯片

图 3 基于MAML的YOLOv3的样本分类结果

Figure 3. Classification results of YOLOv3 based on MAML

下载: 全尺寸图片幻灯片

图 4 YOLOv3模型mAP评估

Figure 4. mAP evaluation of YOLOv3 model

下载: 全尺寸图片幻灯片

图 5 基于MAML的YOLOv3模型mAP评估

Figure 5. mAP evaluation of YOLOv3 model based on MAML

下载: 全尺寸图片幻灯片

图 6 冻结网络层训练的损失函数曲线变化

Figure 6. Change of training loss of frozen network layer

下载: 全尺寸图片幻灯片

图 7 解冻网络层训练的损失函数曲线变化

Figure 7. Change of training loss of thawing network layer

下载: 全尺寸图片幻灯片

图 8 基于MAML的YOLOv3预测结果

Figure 8. Prediction results of YOLOv3 model based on MAML

下载: 全尺寸图片幻灯片

表 1 分类结果的精确度比较结果

Table 1. Accuracy of classification results

Model	Percision/%	mAP/%
YOLOv3	50.7	74.81
YOLOv3 with MAML	56.6	80.05

下载: 导出CSV

参考文献(18)

[1]	FAN Q, ZHUO W, TANG C K, et al. Few-shot object detection with attention-RPN and multi-relation detector[C]//2020 IEEE Conference on Computer Vision and Pattern Recognition. USA: IEEE, 2020: 4012-4021.
[2]	WANG T, ZHANG X P, YUAN L, et al. Few-shot adaptive faster R-CNN[C]//2019 IEEE Conference on Computer Vision and Pattern Recognition. USA: IEEE, 2019: 7166-7175.
[3]	王鹏, 陆振宇, 詹天明, 等. 基于PDSSD改进型神经网络的小目标检测算法[J]. 计算机应用与软件, 2021, 38(1): 149-191.
[4]	FINN C, ABBEEL P, LEVINE S. Model-agnostic meta-learning for fast adaptation of deep networks[C]//Proceedings of the 34th International Conference on Machine Learning. USA: ACM, 2017: 1126-1135.
[5]	李雅倩, 盖成远, 肖存军, 等. 基于细化多尺度深度特征的目标检测网络[J]. 电子学报, 2020, 48(12): 2360-2366.
[6]	REDMON J, FARHADI A. YOLOv3: An incremental improvement[EB/OL]. arXiv. org, (2018-04-08)[2019-04-01]. https://arXiv.org/abs/1804.02767.
[7]	BOCHKOVSKIY A, WANG C Y, LIAO H Y. YOLOv4: Optimal speed and accuracy of object detection[EB/OL]. arXiv. org, (2020-04-23)[2020-09-16]. https://arxiv.org/abs/2004.10934.
[8]	GIRSHICK R, DONAHUE J, DARRELL T, et al. Rich feature hierarchies for accurate object detection and semantic segmentation[C]//2014 IEEE Conference on Computer Vision and Pattern Recognition. USA: IEEE, 2014: 580-587.
[9]	HE K M, ZHANG X Y, REN S Q, et al. Spatial pyramid pooling in deep convolutional networks for visual recognition[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2015, 37(9): 1904-1916. doi: 10.1109/TPAMI.2015.2389824
[10]	GIRSHICK R. Fast R-CNN[C]//2015 IEEE International Conference on Computer Vision. Chile: IEEE, 2015: 1440-1448.
[11]	REN S Q, HE K M, GIRSHICK R, et al. Faster R-CNN: Towards real-time object detection with region proposal networks[C]//2015 International Conference on Neural Information Processing Systems. [s.e.]:[s.n.], 2015: 91-99.
[12]	LIN T, DOLLÁR P, GIRSHICK R, et al. Feature pyramid networks for object detection[C]//2017 IEEE International Conference on Computer Vision. USA: IEEE, 2017: 2117-2125.
[13]	REDMON J, DIVVALA S, GIRSHICK R, et al. You only look once: Unified, real-time object detection[C]//2016 IEEE Conference on Computer Vision and Pattern Recognition. USA: IEEE, 2016: 779-788.
[14]	LIU W, ANGUELOV D, ERHAN D, et al. SSD: Single shot multibox detector[C]//European Conference on Computer Vision. Cham: Springer, 2016: 21-37.
[15]	LIN T, GOYAL P, GIRSHICK R, et al. Focal loss for dense object detection[C]//2017 International Conference on Computer Vision. Italy: IEEE, 2017: 2980-2988.
[16]	TIAN Z, SHEN C H, CHEN H, et al. FCOS: Fully convolutionalone-stage object detection[C]//2019 IEEE International Conference on Computer Vision. Korea: IEEE, 2019: 9626-9635.
[17]	马振伟, 何高奇, 袁玉波. 基于小样本深度学习的通风柜橱窗状态识别方法[J]. 华东理工大学学报(自然科学版), 2020, 46(3): 428-435.
[18]	REDMON J, FARHADI A. YOLO9000: Better, faster, stronger[C]//2017 IEEE Conference on Computer Vision and Pattern Recognition. USA: IEEE, 2017: 7263-7271.