[논문]Voting and Ensemble Schemes Based on CNN Models for Photo-Based Gender Prediction

Jhang, Kyoungson

doi:10.3745/jips.02.0137

Voting and Ensemble Schemes Based on CNN Models for Photo-Based Gender Prediction 원문보기

Journal of information processing systems, v.16 no.4, 2020년, pp.809 - 819

Jhang, Kyoungson (Dept. of Computer Science and Engineering, Chungnam National University)

Abstract ▼ AI-Helper

Gender prediction accuracy increases as convolutional neural network (CNN) architecture evolves. This paper compares voting and ensemble schemes to utilize the already trained five CNN models to further improve gender prediction accuracy. The majority voting usually requires odd-numbered models while the proposed softmax-based voting can utilize any number of models to improve accuracy. The ensemble of CNN models combined with one more fully-connected layer requires further tuning or training of the models combined. With experiments, it is observed that the voting or ensemble of CNN models leads to further improvement of gender prediction accuracy and that especially softmax-based voters always show better gender prediction accuracy than majority voters. Also, compared with softmax-based voters, ensemble models show a slightly better or similar accuracy with added training of the combined CNN models. Softmax-based voting can be a fast and efficient way to get better accuracy without further training since the selection of the top accuracy models among available CNN pre-trained models usually leads to similar accuracy to that of the corresponding ensemble models.

주제어

표/그림 (10)

표 Table 1. VGG architecture [9]
그림 Fig. 1. The skip connection of residual net.
표 Table 2. The residual net architecture summary [14]
그림 Fig. 2. The majority voting of three binary classifier models.
그림 Fig. 3. Softmax-based voting of two models.
그림 Fig. 4. Ensemble of models with a fully-connected layer.
표 Table 3. The statistics of Adience dataset used in the experiment
표 Table 4. Experimental result of CNN models and majority voters
표 Table 5. Experimental result of softmax-based voters
표 Table 6. Experimental result of ensemble voters

AI 본문요약
AI-Helper

* AI 자동 식별 결과로 적합하지 않은 문장이 있을 수 있으니, 이용에 유의하시기 바랍니다.

문제 정의

This paper deals with the accuracy enhancement of photo-based gender prediction utilizing pre-trained CNN models. There have been presented and compared three approaches such as majority voting, softmax-based voting, and ensemble scheme.

가설 설정

(a) The ensemble of two CNN models. (b) A fully-connected layer structure.

제안 방법

(3) Continue to add the next best CNN models to SV_k-1 to construct the best accuracy softmax-based voter SV_k composed of k CNN models.
CNN models used in the experiment are obtained by training for 150 epochs. During training, data augmentation techniques are applied such as random resized crop, random horizontal flip and color jitter of hue, saturation, brightness, and contrast within 20%.
Normally, the 5-fold cross-validation technique[19,20] is used to measure the prediction accuracy as an average value for five cases. However, the experiment aims to test or to show whether the voter or the ensemble method improves the prediction accuracy, rather than to make an accurate comparison with the other methods. Therefore, only one case is used in this experiment as shown in Table 3.
The architecture uses mainly 3×3 convolution (Conv.) filters and consists of five consecutive blocks followed by a classifier layer composed of fully-connected layers.
Ensemble models require further training to get better accuracy. The experiment is performed to find out how many epochs are appropriate for the fine-tuning process. It is observed that 20 epochs of finetuning are enough to get better accuracy and 40 or 100 epochs often lead to worse accuracies.
The model “fc0234” has better accuracy and standard deviation than the model “fc0124” even though the accuracy of the model “1” is better than that of model “3” as shown in Table 4. Therefore, it is necessary to evaluate alternative models one by one to construct the best accuracy model composed of (k+1) CNN models from the best accuracy model composed of k CNN models by the addition of one CNN model.
The first two approaches do not require further training while the last approach requires fine-tuning. With the experiment, it is shown that three approaches helped to achieve better accuracy than the accuracies of the constituent CNN models. Also, it is notable that the softmax-based voter always shows better accuracy than the majority voter does though they consist of the same CNN models.

대상 데이터

As mentioned before, ResNet50, ResNet152, DenseNet, VGG16, VGG19 are used as CNN models for the experiment. As a deep learning platform, PyTorch is used since it provides the aforementioned CNN models and their pretrained weights for ImageNet challenge.
CNN models are initialized with pretrained weights. CNN models used in the experiment are obtained by training for 150 epochs. During training, data augmentation techniques are applied such as random resized crop, random horizontal flip and color jitter of hue, saturation, brightness, and contrast within 20%.
Table 4 also shows accuracies and standard deviations of majority voters. Given five CNN models, the experiment is performed on 10 three-model majority voters and one five-model majority voter. Among three-model voters, “v024” and “v124” show accuracies greater than 95%.
The experimental result with softmax-based voters is shown in Table 5. In the case of softmax-based voters, the experiment is performed on 10 two-model voters, 10 three-model voters, 5 four-model voters, and one five-model voter. In the case of two-model voters, four models such as “sv02”, “sv04”, “sv24”, and “sv34” show accuracies greater than 95%.

성능/효과

Besides, the performance improvement by the softmax-based voting just like other voters or ensemble methods is limited within a certain range, that is, about 1%–2% in this paper.
It is notable that the best two-model softmax-based voter “sv24” shows similar accuracy to those of the best ensemble models “fc024”, “fc0124”, “fc0134”, and “fc01234”.

참고문헌 (20)

E. Eidinger, R. Enbar, and T. Hassner, "Age and gender estimation of unfiltered faces," IEEE Transactions on Information Forensics and Security, vol. 9, no. 12, pp. 2170-2179, 2014.

상세보기
G. Levi and T. Hassner, "Age and gender classification using convolutional neural networks," in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops, Boston, MA, 2015, pp. 34-42.
A. Dehghan, E. G. Ortiz, G. Shu, and S. Z. Masood, "DAGER: deep age, gender and emotion recognition using convolutional neural network," 2017 [Online]. https://arxiv.org/abs/1702.04280.
S. Lapuschkin, A. Binder, K. R. Muller, and W. Samek, "Understanding and comparing deep neural networks for age and gender classification," in Proceedings of the IEEE International Conference on Computer Vision Workshops, Venice, Italy, 2017, pp. 1629-1638.
K. Zhang, C. Gao, L. Guo, M. Sun, X. Yuan, T. X. Han, Z. Zhao, and B. Li, "Age group and gender estimation in the wild with deep RoR architecture," IEEE Access, vol. 5, pp. 22492-22503, 2017.

상세보기
A. Krizhevsky, I. Sutskever, and G. E. Hinton, "ImageNet classification with deep convolutional neural networks," Advances in Neural Information Processing Systems, vol. 25, pp. 1097-1105, 2012.
Y. Jia, E. Shelhamer, J. Donahue, S. Karayev, J. Long, R. Girshick, S. Guadarrama, and T. Darrel, "Caffe: convolutional architecture for fast feature embedding," in Proceedings of the 22nd ACM International Conference on Multimedia, Orlando, FL, 2014, pp. 675-678.
C. Szegedy, W. Liu, Y. Jia, P. Sermanet, S. Reed, D. Anguelov, D. Erhan, V. Vanhoucke, and A. Rabinovich, "Going deeper with convolutions," in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, Boston, MA, 2015, pp. 1-9.
K. Simonyan and A. Zisserman, "Very deep convolutional networks for large-scale image recognition," 2015 [Online]. Available: https://arxiv.org/abs/1409.1556.
O. Russakovsky, J. Deng, H. Su, J. Krause, S. Satheesh, S. Ma, et al., "ImageNet large scale visual recognition challenge," International Journal of Computer Vision, vol. 115, pp. 211-252, 2015.

상세보기
R. Rothe, R. Timofte, and L. Van Gool, "Deep expectation of real and apparent age from a single image without facial landmarks," International Journal of Computer Vision, vol. 126, pp. 144-157, 2018.

상세보기
Wikipedia, "softmas function," 2020 [Online]. Available: https://en.wikipedia.org/wiki/Softmax_function.
Wikipedia, "Convolutional neural network,", 2020 [Online]. Available: https://en.wikipedia.org/wiki/Convolutional_neural_network.
K. He, X. Zhang, S. Ren, and J. Sun, "Deep residual learning for image recognition," 2015 [Online]. Available: https://arxiv.org/abs/1512.03385.
A. Thomas, "An introduction to Global Average Pooling in convolutional neural networks," 2019 [Online]. Available: https://adventuresinmachinelearning.com/global-average-pooling-convolutional-neural-networks/.
G. Huang, Z. Liu, L. Van Der Maaten, and K. Q. Weinberger, "Densely connected convolutional networks," in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, Honolulu, HI, 2017, pp. 4700-4708.
Wikipedia, "Arg max," 2020 [Online]. Available: https://en.wikipedia.org/wiki/Arg_max.
A. Singh, "A comprehensive guide to ensemble learning (with Python codes)," 2018 [Online]. Available: https://www.analyticsvidhya.com/blog/2018/06/comprehensive-guide-for-ensemble-models/.
R. R. Picard and R. D. Cook, "Cross-validation of regression models," Journal of the American Statistical Association, vol. 79, no. 387, pp. 575-583, 1984.

상세보기
S. Arlot and A. Celisse, "A survey of cross-validation procedures for model selection," Statistics Surveys, vol. 4, pp. 40-79, 2010.

상세보기

내보내기 구분	파일저장 인쇄 메일전송
구성항목	기본정보 상세정보 관리번호, 논문명, 저널/프로시딩명, 저자 , 발행년, 권, 호, 시작페이지, 끝페이지, 발행기관 관리번호, 논문명, 대등논문명, 저자 , 저널/프로시딩명, 발행기관, 발행년, 발행언어, 권, 호, 시작페이지, 끝페이지, ISBN, ISSN, 주제분야, 키워드, 초록(한글), 초록(영문), 저자(소속기관)
저장형식	Text(ASCII format) Excel format RefWorks Direct Export RIS format (for Reference Manager, ProCite, EndNote), Scholar's Aids, Mendeley
메일정보	받는사람 (필수) @ 보내는사람 (선택) @ 제목 내용 KISTI 검색결과 이메일 서비스
안내	총 건의 자료가 검색되었습니다. 다운받으실 자료의 인덱스를 입력하세요. (1-10,000) 검색결과의 순서대로 최대 10,000건 까지 다운로드가 가능합니다. 데이타가 많을 경우 속도가 느려질 수 있습니다.(최대 2~3분 소요) 다운로드 파일은 UTF-8 형태로 저장됩니다. 파일의 내용이 제대로 보이지 않을실 때는 웹브라우저 상단의 보기 -> 인코딩 -> 자동선택 여부를 확인하십시오. ~ Text(ASCII format) Excel format

연합인증