Deep Convolution Neural Network Method for Skew Angle Detection in Text Images

Congzhou Guo; Ke Li; Yikun Zhu; Xiaochong Tong; Xiwen Wang

doi:10.3788/LOP202158.1410007

Journals >Laser & Optoelectronics Progress >Volume 58 >Issue 14 >Page 1410007 > Article

Laser & Optoelectronics Progress
Vol. 58, Issue 14, 1410007 (2021)

Deep Convolution Neural Network Method for Skew Angle Detection in Text Images

Congzhou Guo^1、*, Ke Li¹, Yikun Zhu¹, Xiaochong Tong², and Xiwen Wang¹

Author Affiliations

¹Department of Basic, Information Engineering University, Zhengzhou, Henan 450001, China

²School of Surveying and Mapping, Information Engineering University, Zhengzhou, Henan 450001, China

show less

DOI: 10.3788/LOP202158.1410007 Cite this Article Set citation alerts

Congzhou Guo, Ke Li, Yikun Zhu, Xiaochong Tong, Xiwen Wang. Deep Convolution Neural Network Method for Skew Angle Detection in Text Images[J]. Laser & Optoelectronics Progress, 2021, 58(14): 1410007 Copy Citation Text

show less

Fig. 1. DCNN structure of text image tilt angle class detection

Download full size

Fig. 2. Two classification detection of text image skew angle class

Download full size

Fig. 3. The example 1 of text image quality skew angle multi-stage and multi classification detection

Download full size

Fig. 4. The example 2 of text image quality skew angle multi-stage and multi classification detection

Download full size

Fig. 5. The example 3 of text image quality skew angle multi-stage and multi classification detection

Download full size

Fig. 6. Text image simulation data with different skew categories

Download full size

Tilt classification	Tilt angle range	Tilt classification	Tilt angle range
L₁	[0, 30°)	L₇	[0, -30°)
L₂	[30°, 60°)	L₈	[-30°, -60°)
L₃	[60°, 90°)	L₉	[-60°, -90°)
L₄	[90°, 120°)	L₁₀	[-90°, -120°)
L₅	[120°, 150°)	L₁₁	[-120°, -150°)
L₆	[150°, 180°)	L₁₂	[-150°, -180°)

Table 1. Tilt angle class and angle range value

Layer name	Filter size	Stride	Padding
Conv1	3×3×1×64	2	1
Conv2	3×3×64×64	2	1
Max Pooling	2×2	2	0
Conv3	3×3×64×64	2	1
Max Pooling	2×2	2	0
Conv4	3×3×64×64	2	1
Max Pooling	2×2	2	0
Conv5	3×3×64×64	2	1
Conv6	3×3×64×10	2	1

Table 2. Network parameters

Performance index	One stage structure	Multistage structure
Performance index	(Fig.2)	(Fig.3)	(Fig.4)	(Fig.5)
Accuracy	0.982	0.955	0.962	0.956
Recall	0.976	0.966	0.954	0.962
Precision	0.982	0.955	0.962	0.956
F1_Score	0.978	0.965	0.951	0.958
Train time /h	4.321	9.654	9.461	9.622
Test time /ms	7.139	8.112	8.121	8.110

Table 3. Detection results of text image tilt angle classification (12 classifications)

Performance index	Multistage structure
Performance index	(Fig.2’)	(Fig.3’)	(Fig.4’)
Accuracy	0.945	0.943	0.949
Recall	0.946	0.949	0.952
Precision	0.934	0.936	0.937
F1_Score	0.940	0.942	0.944
Train time/h	16.785	16.768	16.642
Test time/ms	13.609	13.434	13.536

Table 4. Detection results of text image tilt angle classification (24 classifications)

Tilt angle classification	Precision of ASTER		Precision of MORAN		Precision of CRNN
Tilt angle classification	Before tilt correction	After tilt correction	Before tilt correction	After tilt correction	Before tilt correction	After tilt correction
L₁	0.966		0.974		0.977
L₂	0.434	0.964	0.454	0.965	0.464	0.960
L₃	0.171	0.961	0.161	0.966	0.181	0.956
L₄	0.205	0.963	0.015	0.953	0.115	0.957
L₅	0	0.959	0	0.959	0	0.959
L₆	0	0.959	0	0.958	0	0.966
L₇	0.807	0.964	0.817	0.947	0.837	0.959
L₈	0.217	0.958	0.247	0.958	0.247	0.951
L₉	0.015	0.946	0.044	0.951	0.055	0.966
L₁₀	0	0.966	0	0.942	0	0.958
L₁₁	0	0.962	0	0.945	0	0.957
L₁₂	0	0.955	0	0.961	0	0.941
L₁₃	0	0.966	0	0.974	0	0.977

Table 5. Comparison of text image recognition before and after skew correction

Congzhou Guo, Ke Li, Yikun Zhu, Xiaochong Tong, Xiwen Wang. Deep Convolution Neural Network Method for Skew Angle Detection in Text Images[J]. Laser & Optoelectronics Progress, 2021, 58(14): 1410007

Download Citation

Set citation alerts for the article

Tools

Set citation alerts for the article

Save the article for my favorites

Paper Information