一种有效的困难样本学习策略

doi:10.19665/j.issn1001-2400.2021.03.013

摘要/Abstract

摘要：

针对困难样本在深度哈希算法中难以收敛以及过多的困难样本产生的噪声干扰问题,提出一种通过损失决定梯度的困难样本学习策略。首先,提出一种非均匀梯度归一化方法,通过计算困难与整体样本损失的比例,对整体样本反向传播梯度进行加权,提高模型对困难样本的学习能力;其次,针对存在过量困难样本的情况,设计了一种加权随机采样方法,根据损失大小对样本进行加权欠采样,滤除噪声并保留少量的困难样本避免过拟合。基于公开数据集,哈希特征检索平均精度值分别约提高了4.7%与3.4%。实验结果表明,该策略改进的哈希算法准确率优于对标哈希算法,能更好地学习到数据集中困难样本的特征信息。

关键词: 采样分析, 梯度算法, 哈希函数, 深度神经网络

Abstract:

To improve the learning efficiency for hard samples and reduce noise interference caused by superfluous hard samples in deep hash algorithm,a generic strategy called Loss to Gradient for hard sample learning is proposed.First,a non-uniform gradient normalization method is proposed to improve the learning ability of models for hard samples.Back propagation gradients are weighted by calculating the loss ratio between hard samples and all samples.Furthermore,a weighted random sampling method is designed for accuracy improvement with superfluous hard samples.According to the loss,training samples are weighted and under-sampled for noise filtering and a small number of hard samples are retained to avoid over-fitting.Based on open datasets,the average accuracy of hash feature retrieval is increased by 4.7% and 3.4%,respectively.Experimental results show that the improved method outperforms other benchmarking methods in accuracy,proving that the feature representation of hard samples in the dataset can be effectively learned.

Key words: sample analysis, gradient method, hash functions, deep neural networks

中图分类号:

TP183

曹艺,蔡晓东. 一种有效的困难样本学习策略[J]. 西安电子科技大学学报, 2021, 48(3): 99-105.

CAO Yi,CAI Xiaodong. Effective learning strategy for hard samples[J]. Journal of Xidian University, 2021, 48(3): 99-105.

图/表 11

图1

图2

图3

表1

图4

图5

表2

表3

图6

表4

表5

参考文献 16

[1]	WANG J D, ZHANG T, SONG J K, et al. A Survey on Learning to Hash[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2018,40(4):769-790. doi: 10.1109/TPAMI.34
[2]	HASHING D P, CAO Z J, SUN Z P, et al. Deep Priority Hashing[C]// Proceedings of the 2018 ACM Multimedia Conference.New York:ACM, 2018: 1653-1661.
[3]	LIU H M, WANG R P, SHAN S G, et al. Deep Supervised Hashing for Fast Image Retrieval[C]// Proceedings of the 2016 IEEE Computer Society Conference on Computer Vision and Pattern Recognition.Piscataway:IEEE, 2016: 2064-2072.
[4]	CAO Z J, LONG M S, WANG J M, et al. HashNet:Deep Learning to Hash by Continuation[C]// Proceedings of the 2017 IEEE International Conference on Computer Vision.Piscataway:IEEE, 2017: 5609-5618.
[5]	LI W J, WANG S, KANG W C. Feature Learning Based Deep Supervised Hashing with Pairwise Labels[C]// Proceedings of the 2016 International Joint Conference Artificial Intelligence.New York:International Joint Conferences on Artificial Intelligence, 2016: 1711-1717.
[6]	ZHU H, LONG M S, WANG J M, et al. Deep Hashing Network for Efficient Similarity Retrieval[C]// Proceedings of the 2016 30th AAAI Conference on Artificial Intelligence.Palo Alto:AAAI, 2016: 2415-2421.
[7]	曾燕, 陈岳林, 蔡晓东. 结合全局与局部池化的深度哈希人脸识别算法[J]. 西安电子科技大学学报, 2018,45(5):163-169.
	ZENG Yan, CHEN Yuelin, CAI Xiaodong. Face Recognition Algorithm for the Deep Hash Combined with Global and Local Pooling[J]. Journal of Xidian University, 2018,45(5):163-169.
[8]	YANG H F, LIN K, CHEN C S. Supervised Learning of Semantics-preserving Hash via Deep Convolutional Neural Networks[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2018,40(2):437-451. doi: 10.1109/TPAMI.2017.2666812
[9]	CAO L F, GAO L L, SONG J K, et al. Multiple Hierarchical Deep Hashing for Large Scale Image Retrieval[J]. Multimed Tools and Application, 2018,77(9):10471-10484. doi: 10.1007/s11042-017-4489-0
[10]	SHI X, SAPKOTA M, XING F, et al. Pairwise Based Deep Ranking Hashing for Histopathology Image Classification and Retrieval[J]. Pattern Recognition, 2018,81:14-22. doi: 10.1016/j.patcog.2018.03.015
[11]	SHRIVASTAVA A, GUPTA A, GIRSHICK R. Training Region-based Object Detectors with Online Hard Example Mining[C]// Proceedings of the 2016 IEEE Computer Society Conference on Computer Vision and Pattern Recognition.Piscataway:IEEE, 2016: 761-769.
[12]	LIN T Y, GOYAL P, GIRSHICK R, et al. Focal Loss for Dense Object Detection[C]// Proceedings of the 2017 IEEE International Conference on Computer Vision.Piscataway:IEEE, 2017: 2999-3007.
[13]	TRAN G S, NGHIEM T P, NGUYEN V T, et al. Improving Accuracy of Lung Nodule Classification Using Deep Learning with Focal Loss[J]. Journal of Healthcare Engineering, 2019,2019(2019):5156416.
[14]	LIU W, CHEN L, CHEN Y. Age Classification Using Convolutional Neural Networks with the Multi-class Focal Loss[C]// Proceedings of the 2018 IOP Conference Series: Materials Science and Engineering.Bristol:IOP Publishing Ltd, 2018: 012043.
[15]	PULGAR F J, RIVERA A J, CHARTE F, et al. On the Impact of Imbalanced Data in Convolutional Neural Networks Performance[C]// Lecture Notes in Computer Science:10334.Heidelberg:Springer Verlag, 2017: 220-232.
[16]	SCHROFF F, KALENICHENKO D, PHILBIN J. FaceNet:a Unified Embedding for Face Recognition and Clustering[C]// Proceedings of the 2015 IEEE Computer Society Conference on Computer Vision and Pattern Recognition.Piscataway:IEEE, 2015: 815-823.

对比方法	CIFAR10				SVHN
对比方法	16 bit	24 bit	36 bit	48 bit	16 bit	24 bit	36 bit	48 bit
文中方法	0.736 8	0.750 4	0.765 0	0.763 9	0.723 4	0.735 8	0.752 3	0.753 7
文献[2]方法	0.715 6	0.756 0	0.775 0	0.770 9	0.696 0	0.725 3	0.740 0	0.745 6
文献[4]方法	0.735 9	0.750 6	0.760 9	0.753 7	0.705 0	0.722 7	0.742 6	0.756 6
文献[5]方法	0.719 6	0.705 2	0.717 5	0.699 3	0.732 0	0.735 0	0.739 0	0.736 0
文献[6]方法	0.715 0	0.703 7	0.719 0	0.716 0	0.722 0	0.732 8	0.734 0	0.720 0

对比方法	CIFAR10
对比方法	16 bit	24 bit	36 bit	48 bit
文中方法	0.734	0.727	0.729	0.731
文献[6]方法	0.715	0.7037	0.719	0.716

对比方法	CIFAR10	MNIST
文中方法	0.731	0.988
Softmax	0.712	0.982

对比方法	CIFAR10
对比方法	16 bit	24 bit	36 bit	48 bit
文中方法	0.734	0.752	0.748	0.752
文献[4]方法*	0.728	0.732	0.736	0.746
文献[6]方法	0.715	0.7037	0.719	0.716

对比方法	平均精确度
加权随机采样	0.752
随机采样	0.734
全部采样简单样本	0.101
全部采样困难样本	0.712
文献[6]方法	0.716