Scene Text Detection Algorithm Based on Color Clustering of Textual Pixels

Fig. 2. Color stratification process. (a) Original image; (b) distribution of all pixels in RGB space; (c) robust textual pixels; (d) distribution of only textual pixels in RGB space; (e) clustering results of textual pixels

Download full size

Fig. 3. Schematic of dangle feature

Download full size

Fig. 4. Comparison of color stratification results before and after grouping. (a) Original image; (b) image before improvement; (c) image after improvement

Download full size

Fig. 5. Flow chart of multi-scale color clustering algorithm

Download full size

Fig. 6. Original images and color stratification results. (a) Images with simple background; (b) images with complex background;(c) images with low contrast

Download full size

Fig. 7. Comparison of color stratification results by different methods. (a) Original image; (b) proposed method; (c) method in Ref. [6]; (d1) result when d=24 in Ref. [10]; (d2) result when d=56 in Ref. [10]

Download full size

Fig. 8. Partial text detection results

Download full size

Fig. 9. Partial images under detection failure. (a) Influence of illumination; (b) text occlusion

Download full size

Table 1. Constraint conditions for merging in text lines

View table

Table 1. Constraint conditions for merging in text lines

Constraint	Definition	Thresholdrange
Height ratio	H_r=height(R₁)/height(R₂)	(0.25,4)
Width ratio	W_r=width(R₁)/width(R₂)	(0.25,4)
Angle	A_angle=arctan(R_h/R_w)	[0,30)
Centroiddistance	d_c= $\begin{matrix} \frac{sqrt (R_{h}^{2} + R_{w}^{2})}{\max [widt h (R_{1}), widt h (R_{2})]} \end{matrix}$	[0,2]
Colordistance	D_color=sqrt $\begin{matrix} [\sum_{C = {r, g, b}} (C_{1} - C_{2})^{2}] \end{matrix}$	[0,30)

Table 2. Performance comparison of different methods on ICDAR2011 database
View table
Table 2. Performance comparison of different methods on ICDAR2011 database
Method R P f
Proposed method 0.71 0.82 0.76
Method in Ref. [10] 0.68 0.82 0.75
Method in Ref. [13] 0.68 0.86 0.76
Method in Ref. [14] 0.69 0.81 0.75
Method in Ref. [15] 0.65 0.84 0.73

Table 3. Performance comparison of different methods on ICDAR2013 database
View table
Table 3. Performance comparison of different methods on ICDAR2013 database
Method R P f
Proposed method 0.73 0.82 0.77
Method in Ref. [10] 0.70 0.84 0.76
Method in Ref. [16] 0.65 0.84 0.73
Method in Ref. [17] 0.68 0.79 0.73
Method in Ref. [18] 0.63 0.85 0.72

Tools

Get Citation

Copy Citation Text

Min Li, Jianbin Zheng, Enqi Zhan, Yang Wang. Scene Text Detection Algorithm Based on Color Clustering of Textual Pixels[J]. Laser & Optoelectronics Progress, 2019, 56(7): 071006

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites