Unsegmented Recognition of Handwritten Numerical Strings Based on Mask-RCNN

Zhiyong Tao; Yueming Han; Sen Lin

doi:10.3788/LOP57.141012

Journals >Laser & Optoelectronics Progress >Volume 57 >Issue 14 >Page 141012 > Article

Laser & Optoelectronics Progress
Vol. 57, Issue 14, 141012 (2020)

Unsegmented Recognition of Handwritten Numerical Strings Based on Mask-RCNN

Zhiyong Tao¹, Yueming Han^1、2、*, and Sen Lin¹

Author Affiliations

¹School of Electronic & Information Engineering, Liaoning Technical University, Huludao, Liaoning 125105, China;

²Fuxin Lixing Technology Company Limited, Fuxin, Liaoning 123000, China

show less

DOI: 10.3788/LOP57.141012 Cite this Article Set citation alerts

Zhiyong Tao, Yueming Han, Sen Lin. Unsegmented Recognition of Handwritten Numerical Strings Based on Mask-RCNN[J]. Laser & Optoelectronics Progress, 2020, 57(14): 141012 Copy Citation Text

show less

Fig. 1. Mask-RCNN structure

Download full size

Fig. 2. RPN working mechanism

Download full size

Fig. 3. Schematic principle of ROI Align layer algorithm

Download full size

Fig. 4. Isolated number set

Download full size

Fig. 5. Multi-digit string set

Download full size

Fig. 6. Loss change curves. (a) Total network loss; (b) frame regression loss; (c) mask loss; (d) classification loss

Download full size

Type	1-digit	2-digits	3-digits	4-digits	5-digits	6-digits
Training	2240	1680	1120	560	560	560
Test	960	720	480	240	240	240

Table 1. Data amount distribution of training dataset and test dataset

Parameter	Value
Number of training datasets	6720
Number of test datasets	2880
Step per epoch	30
Batch	224
Epochλ	3005×10^-4

Table 2. Experimental parameters setting

Type	1-digit	2-digits	3-digits	4-digits	5-digits	6-digits
Accuracy rate /%	99.9	98.9	98.5	97.1	96.3	95.9

Table 3. Mask-RCNN accuracy rate on test dataset

Method	1-digit	2-digits	3-digits	4-digits	5-digits	6-digits	Overall
Ref. [20]	-	94.8	91.6	91.3	88.3	89.0	91.0
Ref. [21]	-	97.6	96.2	94.2	94.0	93.8	95.2
Ref. [22]	-	98.9	97.2	96.1	95.8	96.1	96.8
Ref. [23]	-	99.0	97.3	96.5	95.9	96.6	97.1
Ref. [6]	99.6	97.6	96.2	94.6	-	-	97.0
Proposed	99.9	98.9	98.5	97.1	96.3	95.9	97.8

Table 4. Comparison of different identification methods on NIST SD19 datasetunit: %

Zhiyong Tao, Yueming Han, Sen Lin. Unsegmented Recognition of Handwritten Numerical Strings Based on Mask-RCNN[J]. Laser & Optoelectronics Progress, 2020, 57(14): 141012

Download Citation

Set citation alerts for the article

Tools

Set citation alerts for the article

Save the article for my favorites

Paper Information