基于自適應(yīng)權(quán)值裁剪的Adaboost快速訓(xùn)練算法

余陸斌; 杜啟亮; 田聯(lián)房

doi:10.11999/JEIT190473

基于自適應(yīng)權(quán)值裁剪的Adaboost快速訓(xùn)練算法

doi: 10.11999/JEIT190473 cstr: 32379.14.JEIT190473

華南理工大學(xué)自動化科學(xué)與工程學(xué)院廣州 510640

基金項目: 海防公益類項目(201505002)，廣東省重點研發(fā)計劃-新一代人工智能(20180109)，廣州市產(chǎn)業(yè)技術(shù)重大攻關(guān)計劃(2019-01-01-12-1006-0001)，廣東省科學(xué)技術(shù)廳重大科技計劃項目(2016B090912001)，中央高校基本科研業(yè)務(wù)費專項資金(2018KZ05)

詳細(xì)信息

作者簡介:
余陸斌：男，1994年生，博士生，主要研究方向為機器學(xué)習(xí)、機器視覺

杜啟亮：男，1980年生，副研究員，博士，主要研究方向為機器人、機器視覺

田聯(lián)房：男，1968年生，教授，博士，主要研究方向為模式識別、人工智能

通訊作者:
杜啟亮　qldu@scut.edu.cn

中圖分類號: TP391
計量
- 文章訪問數(shù): 1183
- HTML全文瀏覽量: 428
- PDF下載量: 62
- 被引次數(shù): 0
出版歷程
- 收稿日期: 2019-06-27
- 修回日期: 2020-04-19
- 網(wǎng)絡(luò)出版日期: 2020-08-31
- 刊出日期: 2020-11-16

Fast Training Adaboost Algorithm Based on Adaptive Weight Trimming

College of Automation Science and Engineering, South China University of Technology, Guangzhou 510640, China

Funds: The Coast defence Public Welfare Project (201505002), Guangdong Province Key R&D Program-A New Generation of Artificial Intelligence (20180109), Guangzhou City Industrial Technology Major Research Project (2019-01-01-12-1006-0001), The Major Science and Technology Plan Project of Guangdong Science and Technology Department (2016B090912001), The Special Fund for Basic Scientific Research in Central Colleges and Universities (2018KZ05)

摘要

摘要: Adaboost是一種廣泛使用的機器學(xué)習(xí)算法，然而Adaboost算法在訓(xùn)練時耗時十分嚴(yán)重。針對該問題，該文提出一種基于自適應(yīng)權(quán)值的Adaboost快速訓(xùn)練算法AWTAdaboost。該算法首先統(tǒng)計每一輪迭代的樣本權(quán)值分布，再結(jié)合當(dāng)前樣本權(quán)值的最大值和樣本集規(guī)模計算出裁剪系數(shù)，權(quán)值小于裁剪系數(shù)的樣本將不參與訓(xùn)練，進(jìn)而加快了訓(xùn)練速度。在INRIA數(shù)據(jù)集和自定義數(shù)據(jù)集上的實驗表明，該文算法能在保證檢測效果的情況下大幅加快訓(xùn)練速度，相比于其他快速訓(xùn)練算法，在訓(xùn)練時間接近的情況下有更好的檢測效果。
- 目標(biāo)檢測 /
- Adaboost算法 /
- 快速訓(xùn)練 /
- 自適應(yīng) /
- 權(quán)值分布
Abstract: The Adaboost algorithm provides noteworthy benefits over the traditional machine algorithms for numerous applications, including face recognition, text recognition, and pedestrian detection. However, it takes a lot of time during the training process that affects the overall performance. Adaboost fast training algorithm based on adaptive weight (Adaptable Weight Trimming Adaboost, AWTAdaboost) is proposed in this work to address the aforementioned issue. First, the algorithm counts the current sample weight distribution of each iteration. Then, it combines the maximum value of current sample weights with data size to calculate the adaptable coefficients. The sample whose weight is less than the adaptable coefficients is discarded, that speeds up the training. The experimental results validate that it can significantly speed up the training speed while ensuring the detection effect. Compared with other fast training algorithms, the detection effect is better when the training time is close to each other.
- Object detection /
- Adaboost algorithm /
- Fast traing /
- Adaptive /
- Weight distribution

HTML全文

圖 1 自定義數(shù)據(jù)集樣本示例

下載: 全尺寸圖片幻燈片

圖 2 各算法在INRIA數(shù)據(jù)集上的錯誤率

下載: 全尺寸圖片幻燈片

圖 3 各算法在自定義數(shù)據(jù)集上的錯誤率

下載: 全尺寸圖片幻燈片

圖 4 各算法的訓(xùn)練時間

下載: 全尺寸圖片幻燈片

圖 5 AWTAdaboost算法在訓(xùn)練時保留樣本比例

下載: 全尺寸圖片幻燈片

表 1 各算法在兩個數(shù)據(jù)集上的錯誤率

	INRIA數(shù)據(jù)集		自定義數(shù)據(jù)集
	訓(xùn)練集錯誤率	測試集錯誤率	訓(xùn)練集錯誤率	測試集錯誤率
Adaboost	0.0000	0.0285	0.0000	0.0296
SWTAdaboost	0.0395	0.0768	0.0538	0.1089
DWTAdaboost	0.0000	0.0466	0.0194	0.0735
WNS-Adaboost	0.0000	0.0356	0.0006	0.0439
GAdaboost	0.0563	0.1108	0.0724	0.1345
PCA+DRAdaboost	0.0000	0.0413	0.0000	0.0539
AWTAdaboost	0.0000	0.0302	0.0000	0.0324

下載: 導(dǎo)出CSV

表 2 各算法訓(xùn)練時間對比

算法	INRIA數(shù)據(jù)集相對訓(xùn)練時間	自定義數(shù)據(jù)集相對訓(xùn)練時間
Adaboost	1.0000	1.0000
SWTAdaboost	0.6237	0.6547
DWTAdaboost	0.6347	0.6551
WNS-Adaboost	0.5814	0.5919
GAdaboost	0.4482	0.4636
PCA+DRAdaboost	0.5124	0.5324
AWTAdaboost	0.5570	0.5732
注：表中只記錄了SWTAdaboost提前停止迭代前的訓(xùn)練時間和相同$\beta $下DWTAdaboost的訓(xùn)練時間。

下載: 導(dǎo)出CSV

參考文獻(xiàn)(15)

VALIANT L G. A theory of the learnable[C]. The 16th Annual ACM Symposium on Theory of Computing, New York, USA, 1984: 436–445.

KEARNS M and VALIANT L. Cryptographic limitations on learning Boolean formulae and finite automata[J]. Journal of the ACM, 1994, 41(1): 67–95. doi: 10.1145/174644.174647

SCHAPIRE R E. The strength of weak learnability[J]. Machine Learning, 1990, 5(2): 197–227.

FREUND Y and SCHAPIRE R E. A decision-theoretic generalization of on-line learning and an application to boosting[J]. Journal of Computer and System Sciences, 1997, 55(1): 119–139. doi: 10.1006/jcss.1997.1504

FREUND Y and SCHAPIRE R E. Experiments with a new boosting algorithm[C]. International Conference on Machine Learning, Bari, Italy, 1996: 148–156.

ZHANG Xingqiang and DING Jiajun. An improved Adaboost face detection algorithm based on the different sample weights[C]. The 20th IEEE International Conference on Computer Supported Cooperative Work in Design, Nanchang, China, 2016: 436–439.

CHO H, SUNG M, and JUN B. Canny text detector: Fast and robust scene text localization algorithm[C]. 2016 IEEE Conference on Computer Vision and Pattern Recognition, Las Vegas, USA, 2016: 3566–3573.

GAO Chenqiang, LI Pei, ZHANG Yajun, et al. People counting based on head detection combining Adaboost and CNN in crowded surveillance environment[J]. Neurocomputing, 2016, 208: 108–116. doi: 10.1016/j.neucom.2016.01.097

FRIEDMAN J, HASTIE T, and TIBSHIRANI R. Additive logistic regression: A statistical view of boosting (With discussion and a rejoinder by the authors)[J]. Annals of Statistics, 2000, 28(2): 337–407.

賈慧星, 章毓晉. 基于動態(tài)權(quán)重裁剪的快速Adaboost訓(xùn)練算法[J]. 計算機學(xué)報, 2009, 32(2): 336–341. doi: 10.3724/SP.J.1016.2009.00336

JIA Huixing and ZHANG Yujin. Fast Adaboost training algorithm by dynamic weight trimming[J]. Chinese Journal of Computers, 2009, 32(2): 336–341. doi: 10.3724/SP.J.1016.2009.00336

SEYEDHOSSEINI M, PAIVA A R C, and TASDIZEN T. Fast AdaBoost training using weighted novelty selection[C]. 2011 International Joint Conference on Neural Networks, San Jose, USA, 2011: 1245–1250.

TOLBA M F and MOUSTAFA M. GAdaBoost: Accelerating adaboost feature selection with genetic algorithms[C]. The 8th International Joint Conference on Computational Intelligence, Porto, Portugal, 2016: 156-163.

YUAN Shuang and Lü Cixing. Fast adaboost algorithm based on weight constraints[C]. 2015 IEEE International Conference on Cyber Technology in Automation, Control, and Intelligent Systems, Shenyang, China, 2015: 825–828.

袁雙, 呂賜興. 基于PCA改進(jìn)的快速Adaboost算法研究[J]. 科學(xué)技術(shù)與工程, 2015, 15(29): 62–66. doi: 10.3969/j.issn.1671-1815.2015.29.011

YUAN Shuang and Lü Cixing. Fast adaboost algorithm based on improved PCA[J]. Science Technology and Engineering, 2015, 15(29): 62–66. doi: 10.3969/j.issn.1671-1815.2015.29.011

DALAL N and TRIGGS B. Histograms of oriented gradients for human detection[C]. 2005 IEEE Computer Society Conference on Computer Vision and Pattern Recognition, San Diego, USA, 2005: 886–893.

相關(guān)文章

施引文獻(xiàn)

資源附件(0)

訪問統(tǒng)計