基于改進(jìn)深度卷積神經(jīng)網(wǎng)絡(luò)的紙幣識(shí)別研究

蓋杉; 鮑中運(yùn)

doi:10.11999/JEIT181097

基于改進(jìn)深度卷積神經(jīng)網(wǎng)絡(luò)的紙幣識(shí)別研究

doi: 10.11999/JEIT181097

蓋杉^,,
鮑中運(yùn)

南昌航空大學(xué)信息工程學(xué)院南昌 330063

基金項(xiàng)目: 國(guó)家自然科學(xué)基金(61563037)，江西省杰出青年計(jì)劃(20171BCB23057)

詳細(xì)信息

作者簡(jiǎn)介:
蓋杉：男，1980年生，副教授，碩士生導(dǎo)師，研究方向?yàn)橛?jì)算機(jī)視覺、圖像處理、深度學(xué)習(xí)

鮑中運(yùn)：男，1990年生，碩士生，研究方向?yàn)橛?jì)算機(jī)視覺、圖像處理、深度學(xué)習(xí)

通訊作者:
蓋杉　gaishan@nchu.edu.cn

中圖分類號(hào): TP391.41; TP181
計(jì)量
- 文章訪問(wèn)數(shù): 4230
- HTML全文瀏覽量: 1281
- PDF下載量: 116
- 被引次數(shù): 0
出版歷程
- 收稿日期: 2018-11-28
- 修回日期: 2019-03-27
- 網(wǎng)絡(luò)出版日期: 2019-04-21
- 刊出日期: 2019-08-01

Banknote Recognition Research Based on Improved Deep Convolutional Neural Network

Shan GAI^,,
Zhongyun BAO

School ofInformation Engineering, Nanchang HangkongUniversity, Nanchang 330063, China

Funds: The National Natural Science Foundation of China(61563037), The Outstanding Youth Scheme of Jiangxi Province (20171BCB23057)

摘要

摘要: 針對(duì)如何提高紙幣識(shí)別率的問(wèn)題，該文提出一種改進(jìn)深度卷積神經(jīng)網(wǎng)絡(luò)(DCNN)的紙幣識(shí)別算法。該算法首先通過(guò)融合遷移學(xué)習(xí)、帶泄露整流(Leaky ReLU)函數(shù)、批量歸一化(BN)和多層次殘差單元構(gòu)造深度卷積層，對(duì)輸入的不同尺寸紙幣進(jìn)行穩(wěn)定而快速的特征提取與學(xué)習(xí)；然后采用改進(jìn)的多層次空間金字塔池化算法對(duì)提取的紙幣特征實(shí)現(xiàn)固定大小的輸出表示；最后通過(guò)網(wǎng)絡(luò)全連接層和softmax層實(shí)現(xiàn)紙幣圖像分類。實(shí)驗(yàn)結(jié)果表明，該算法在分類性能、泛化能力與穩(wěn)定性上明顯優(yōu)于常用的紙幣分類算法；同時(shí)該算法也能夠滿足紙幣清分系統(tǒng)的實(shí)時(shí)性要求。
- 紙幣識(shí)別 /
- 深度卷積神經(jīng)網(wǎng)絡(luò) /
- 殘差學(xué)習(xí) /
- 空間金字塔池化
Abstract: In order to improve the recognition rate of banknotes, the improved banknote recognition algorithm based on Deep Convolutional Neural Network(DCNN) is proposed. Firstly, the algorithm constructs a deep convolution layer by integrating transfer learning, Leaky-Rectified Liner Unit (Leaky ReLU) function, Batch Normalization(BN) and multi-level residual unit that perform stable and fast feature extraction and learning on input different size banknotes. Secondly, a fixed-size output representation of the extracted banknote features is obtained by using the improved multi-level spatial pyramid pooling algorithm. Finally, the banknote classification is implemented by the full connection layer and the softmax layer of the network. The experimental results show that the proposed algorithm can effectively improve the recognition rate of banknotes, and has better generalization ability and robustness than the traditional banknote classification method. Meanwhile, the algorithm can meet the real-time requirements of the banknote sorting system.
- Banknote recognition /
- Deep Convolutional Neural Network (DCNN) /
- Residual learning /
- Spatial pyramid pooling

HTML全文

圖 1 算法結(jié)構(gòu)示意圖

下載: 全尺寸圖片幻燈片

圖 2 紙幣圖像預(yù)處理(RMB-100)

下載: 全尺寸圖片幻燈片

圖 3 紙幣圖像預(yù)處理(USD-100)

下載: 全尺寸圖片幻燈片

圖 4 紙幣圖像預(yù)處理(EUR-500)

下載: 全尺寸圖片幻燈片

圖 5 多層次殘差單元結(jié)構(gòu)圖

下載: 全尺寸圖片幻燈片

圖 6 多層次空間金字塔池化算法結(jié)構(gòu)框架

下載: 全尺寸圖片幻燈片

圖 7 紙幣圖像的4個(gè)面向

下載: 全尺寸圖片幻燈片

表 1 紙幣數(shù)據(jù)庫(kù)

紙幣種類	紙幣面值	紙幣分類	紙幣樣本數(shù)	訓(xùn)練樣本數(shù)	測(cè)試樣本數(shù)
人民幣(RMB)	5, 10, 20, 50, 100	20	46000	36000	10000
美元(USD)	1, 2, 10, 20, 50, 100	24	38000	25000	13000
歐元(EUR)	5, 10, 20, 50, 100, 200, 500	28	35000	26000	9000

下載: 導(dǎo)出CSV

表 2 數(shù)據(jù)庫(kù)DB1平均識(shí)別率(%)

人民幣	網(wǎng)格特征^[3]	自由掩模^[2]	VGGNet19^[10]	PReLU-net^18]	BN-inception^[16]	ResNet-34B^[13]	本文算法
100	74.25	76.44	91.52	91.45	92.38	94.16	96.68
50	74.02	74.75	90.83	91.76	92.11	95.98	97.80
20	75.23	76.88	92.34	91.56	93.64	94.88	95.03
10	80.12	83.34	94.06	94.76	95.67	96.86	96.97
5	83.24	80.57	93.16	93.27	95.53	95.66	97.82

下載: 導(dǎo)出CSV

表 3 數(shù)據(jù)庫(kù)DB2平均識(shí)別率(%)

美元	網(wǎng)格特征^[3]	自由掩模^[2]	VGGNet19^[10]	PReLU-net^[18]	BN-inception^[16]	ResNet-34B^[13]	本文算法
100	70.13	72.24	89.26	91.33	93.25	94.46	95.67
50	73.14	72.28	91.35	91.49	92.98	94.29	94.96
20	74.56	77.82	90.23	92.14	93.05	95.11	95.89
10	76.21	75.34	91.25	93.34	93.67	94.28	95.15
2	78.11	80.12	92.13	92.86	93.58	95.67	96.75
1	81.23	80.02	91.24	90.36	94.27	96.16	97.98

下載: 導(dǎo)出CSV

表 4 數(shù)據(jù)庫(kù)DB3平均識(shí)別率(%)

歐元	網(wǎng)格特征^[3]	自由掩模^[2]	VGGNet19^[10]	PReLU-net^[18]	BN-inception^[16]	ResNet-34B^[13]	本文算法
500	81.12	84.23	93.25	92.91	94.56	94.93	96.98
200	81.65	82.32	93.24	94.13	94.68	95.12	98.20
100	85.46	86.94	94.12	94.67	95.23	96.11	97.75
50	79.25	83.24	93.20	93.12	94.35	95, 29	96.79
20	83.24	84.52	94.25	95.28	95.64	96.33	98.76
10	85.33	87.12	94.24	94.76	94.19	97.20	97.88
5	84.20	83.52	94.16	93.26	95.12	95.78	97.89

下載: 導(dǎo)出CSV

表 5 污損紙幣實(shí)際測(cè)試識(shí)別率(%)

污損樣本	網(wǎng)格特征^[3]	自由掩模^[2]	VGGNet19^[10]	PReLU-net^[18]	BN-inception^[16]	ResNet-34B^[13]	本文算法
DB1(16100)	78.65	82.49	92.45	93.18	94.37	95.06	97.58
DB2(15960)	75.42	79.16	88.24	91.07	92.53	94.84	96.75
DB3(10500)	80.28	83.17	91.52	93.65	95.18	96.78	97.29

下載: 導(dǎo)出CSV

表 6 不同識(shí)別算法運(yùn)行時(shí)間(s)

自由掩模^[2]	網(wǎng)格特征^[3]	VGGNet19^[10]	PReLU-Net^[18]	BN-inception^[16]	ResNet-34B^[13]	本文算法
0.98	0.85	1.97	1.72	1.58	1.24	1.06

下載: 導(dǎo)出CSV

參考文獻(xiàn)(18)

KATO N, SUZUKI M, OMACHI S, et al. A handwritten character recognition system using directional element feature and asymmetric mahalanobis distance[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 1999, 21(3): 258–262. doi: 10.1109/34.754617

TAKEDA F and OMATU S. High speed paper currency recognition by neural networks[J]. IEEE Transactions on Neural Networks, 1995, 6(1): 73–77. doi: 10.1109/72.363448

劉家鋒, 劉松波, 唐降龍. 一種實(shí)時(shí)紙幣識(shí)別方法的研究[J]. 計(jì)算機(jī)研究與發(fā)展, 2003, 40(7): 1057–1061.

LIU Jiafeng, LIU Songbo, and TANG Xianglong. An algorithm of real-time paper currency recongnition[J]. Journal of Computer Research and Development, 2003, 40(7): 1057–1061.

CHOI E, LEE J, and YOON J. Feature extraction for bank note classification using wavelet transform[C]. The IEEE 18th International Conference on Pattern Recognition (ICPR), Hong Kong, China, 2006: 934–937. doi: 10.1109/ICPR.2006.553.

GAI Shan, YANG Guowei, and WAN Minghua. Employing quaternion wavelet transform for banknote classification[J]. Neurocomputing, 2013, 118: 171–178. doi: 10.1016/j.neucom.2013.02.029

JIN Ye, SONG Ling, TANG Xianglong, et al. A hierarchical approach for banknote image processing using homogeneity and FFD model[J]. IEEE Signal Processing Letters, 2008, 15: 425–428. doi: 10.1109/LSP.2008.921470

吳震東, 王雅妮, 章堅(jiān)武. 基于深度學(xué)習(xí)的污損指紋識(shí)別研究[J]. 電子與信息學(xué)報(bào), 2017, 39(7): 1585–1591. doi: 10.11999/JEIT161121

WU Zhendong, WANG Yani, and ZHANG Jianwu. Fouling and damaged fingerprint recognition based on deep learning[J]. Journal of Electronics &Information Technology, 2017, 39(7): 1585–1591. doi: 10.11999/JEIT161121

樊養(yǎng)余, 李祖賀, 王鳳琴, 等. 基于跨領(lǐng)域卷積稀疏自動(dòng)編碼器的抽象圖像情緒性分類[J]. 電子與信息學(xué)報(bào), 2017, 39(1): 167–175. doi: 10.11999/JEIT160241

FAN Yangyu, LI Zuhe, WANG Fengqin, et al. Affective abstract image classification based on convolutional sparse autoencoders across different domains[J]. Journal of Electronics &Information Technology, 2017, 39(1): 167–175. doi: 10.11999/JEIT160241

KRIZHEVSKY A, SUTSKEVER I, and HINTON G E. ImageNet classification with deep convolutional neural networks[C]. The 25th International Conference on Neural Information Processing Systems, Nevada, USA, 2012: 1097–1105.

SIMONYAN K and ZISSERMAN A. Very deep convolutional networks for large-scale image recognition[C]. International Conference on Learning Representations (ICLR), Banff, Canada, 2015: 168–175.

SZEGEDY C, LIU Wei, JIA Yangqing, et al. Going deeper with convolutions[C]. 2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Boston, USA, 2015: 1–9. doi: 10.1109/CVPR.2015.7298594.

SZEGEDY C, VANHOUCKE V, IOFFE S, et al. Rethinking the inception architecture for computer vision[C]. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Las Vegas, USA, 2016: 2818–2826. doi: 10.1109/CVPR.2016.308.

HE Kaiming, ZHANG Xiangyu, REN Shaoqing, et al. Deep residual learning for image recognition[C]. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), Las Vegas, USA, 2016: 770–778. doi: 10.1109/CVPR.2016.90.

HE Kaiming, ZHANG Xiangyu, REN Shaoqing, 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

PENG Peixi, TIAN Yonghong, XIANG Tao, et al. Joint semantic and latent attribute modeling for cross-class transfer learning[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2018, 40(7): 1625–1638. doi: 10.1109/TPAMI.2017.2723882

IOFFE S and SZEGEDY C. Batch normalization: accelerating deep network training by reducing internal covariate shift[C]. Proceedings of the 32nd International Conference on International Conference on Machine Learning, Lille, France, 2015: 448–456.

KINGMA D P and BA J. Adam: A method for stochastic optimization[C]. Proceedings of the 3rd International Conference on Learning Representations (ICLR), San Diego, USA, 2015: 1–8.

HE Kaiming, ZHANG Xiangyu, REN Shaoqing, et al. Delving deep into rectifiers: Surpassing human-level performance on ImageNet classification[C]. 2015 IEEE International Conference on Computer Vision (ICCV), Santiago, Chile, 2015: 1026–1034. doi: 10.1109/ICCV.2015.123.

相關(guān)文章

施引文獻(xiàn)

資源附件(0)

訪問(wèn)統(tǒng)計(jì)