GAPF: Curve Text Detection based on Generative Adversarial Networks

and Pixel Fluctuations

Jun Yang

, Zhaogong Zhang

and Xuexia Wang

School of Computer Science and Technology, Heilongjiang University, Harbin, China

Department of Mathematics, University of North Texas, Denton, U.S.A.

Keywords:

Deep Learning, Pattern Recognition, Curved Text Detection, Generative Adversarial Networks, Pixel

Fluctuations Numbers.

Abstract:

Scene text detection has witnessed rapid progress especially with the recent development of convolutional

neural networks. However, curved text detection is still a difﬁcult problem that has not been addressed sufﬁ-

ciently. Presently, the most advanced method is based on segmentation to detect curved text. However, most

segmentation algorithms based on convolutional neural networks have the problem of inaccurate segmentation

results. In order to improve the effect of image segmentation, we propose a semantic segmentation network

model based on generative adversarial networks and pixel ﬂuctuations, denoted as GAPF; which is able to

effectively improve the accuracy of text segmentation. The model consists of two parts: the generative model

and the discriminative model. The main function of the generative model is to generate semantic segmentation

graph, and then the discriminative model and generative model perform adversarial learning, which optimize

the generative model to make the generated image closer to the ground truth. In this paper, the information

about pixel ﬂuctuations numbers is input into the generative network as the segmentation condition to enhance

the invariance of translation and rotation. Finally, a text boundary generation algorithm for text is designed,

and the ﬁnal detection result is obtained from the segmentation result. Experimental results on CTW1500,

Total-Text, ICDAR 2015 and MSRA-TD500 demonstrate the effectiveness of our work.

1 INTRODUCTION

Scene text is one of the most common objects in im-

ages, which usually appears on license plates, prod-

uct packages, billboards, and carries rich semantic

information. Scene text detection is to identify text

regions of the given scene text images, which is

also an important prerequisite for many multimedia

tasks, such as image understanding and video analy-

sis. Compared to common objects, scene text is born

with multiple orientations, large aspect ratios, arbi-

trary shapes or layouts, and complex backgrounds;

which creates difﬁculties for detection and recogni-

tion. With the development of convolutional neural

networks(Kaiming,2016), there have been many at-

tempts on text detection in natural scenes and great

progress has been achieved in recent years. The early

attempts to detect text are with annotations of hori-

zontal texts and the approaches for arbitrary-oriented

scene text detection have been also proposed. How-

ever, in real-world scenarios, there are many text re-

gions with irregular shapes, such as curve words or

logos. It is very challenging to detect these regions

with different shapes.

The detection of curve text and arbitrary shape

text is almost always based on semantic segmentation,

and the ﬁnal detection results are obtained from the

segmentation graph through the post-processing algo-

rithm. From this perspective, accurate segmentation

results are an important prerequisite for improving the

accuracy of text detection. However, most semantic

segmentation algorithms based on convolutional neu-

ral networks have inaccurate segmentation. Gener-

ative adversarial networks(GAN)(Goodfellow,2014)

have been proven to effectively improve network per-

formance. Based on the idea of GAN, this paper de-

signs a semantic segmentation network model based

on generative adversarial learning to generate accu-

rate segmentation results for text images. Finally, the

ﬁnal text detection results are obtained by the text

boundary generation algorithm.

The contributions of this paper can be summarized

as follows.

Yang, J., Zhang, Z. and Wang, X.

GAPF: Curve Text Detection based on Generative Adversarial Networks and Pixel Fluctuations.

DOI: 10.5220/0010298905450552

In Proceedings of the 16th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP 2021) - Volume 5: VISAPP, pages

545-552

ISBN: 978-989-758-488-6

545

• The generative adversarial network is introduced

into the ﬁeld of text detection. Combining with

the idea of conditional generative adversarial net-

works, the original image is used as the input of

the generator to generate the desired semantic seg-

mentation results. By training the generator so

that the discriminator cannot distinguish between

the image generated by the generator and the re-

sults of manual annotation, the generator can gen-

erate satisfactory image segmentation results.

• Information on the pixel ﬂuctuations numbers is

proposed. The pixel ﬂuctuations numbers as a

condition are input of the generator network. This

paper ﬁrst calculates the pixel ﬂuctuations num-

bers of each pixel in its pixel interval, then stacks

the calculation result with the original picture into

the generator.

• Lastly, a text boundary generation algorithm is de-

signed. The bounding box of the text is generated

from the segmentation result, and the ﬁnal output

is obtained.

2 RELATED WORK

Scene text detection based on deep learning meth-

ods has achieved remarkable results over the past

few years. Modern methods are mostly based on

deep neural networks, which can be coarsely classi-

ﬁed into two categories: regression-based methods

and segmentation-based methods.

Regression-based methods mainly draw inspira-

tion from general object detection frameworks. Based

on Faster-RCNN(Ren,2017), Ma et al. (Ma,2018) de-

vised Rotation Region Proposal Networks (RRPN)

to detect arbitrary Oriented text in natural images.

Textboxes (Liao,2017)adopted SSD (Liu,2016) and

added long default boxes and ﬁlters to handle the sig-

niﬁcant variation of aspect ratios of text instances.

EAST (Zhou,2017) uses a single neural network to

directly predict score map, rotation angle and text

boxes for each pixel. Tian et al. introduced CTPN

(Zhi,2016) using LSTM (Alex,2005) to link several

text proposals. These methods can handle multi-

directional text but may have shortcomings when

dealing with curved text, which widely exist in real-

world scenarios.

Segmentation-based methods are mainly in-

spired by fully convolutional networks (FCN)

(Jonathan,2015). Zhang et al. (Zheng,2016) ﬁrst

adopted FCN to extract text blocks and detect charac-

ter candidates from those text blocks via MSER. Pix-

elLink (Dan,2018) separated texts lying close to each

other by predicting pixel connections between differ-

ent text. The framework of TextSnake (Long,2018)

considered a text instance as a sequence of ordered

disks. To deal with the problem of separation of the

close text instances, Li et al. (Li,2019) designed the

PSENet, a progressive scale algorithm to gradually

expand the predeﬁned kernels. The methods reviewed

above have made signiﬁcant improvements on various

benchmarks in this ﬁeld. However, most segmenta-

tion algorithms fail to produce accurate segmentation

results, resulting in low accuracy of the ﬁnal detection

results.

3 PROPOSED METHOD

3.1 Pixel Fluctuations of Grayscale

Image

We know that images are made up of pixels. For ex-

ample, Figure 1(a) is a grayscale image. If we were

to draw the pixel values of the twentieth line into a

curve, we would get the following Figure 1(b).

(a)

(b)

Figure 1: Visualization of pixel ﬂuctuations. (a) is the

grayscale image of the original image. (b) is the ﬂuctua-

tion curve of the twentieth line of the grayscale image.

It can be seen that the curve ﬂuctuates up and

down continuously, some areas have relatively small

ﬂuctuations, and some areas suddenly show large

ﬂuctuations. By comparing the images, we can see

that the curve ﬂuctuates wildly in the text line area,

while the ﬂuctuation is relatively smooth in the back-

ground area. This shows that the ﬂuctuation is closely

VISAPP 2021 - 16th International Conference on Computer Vision Theory and Applications

546

related to the image. Large ﬂuctuations represent

sharp changes in color; small ﬂuctuations represent

smooth color transitions. But then the question be-

comes how to quantify this ﬂuctuation. We propose

to use pixel ﬂuctuations numbers of the image, which

can ﬁnd out whether there is a ﬂuctuation in the pixel

interval. First, we need to deﬁne the total variation.

3.1.1 Total Variation

Let f (x) be the function deﬁned on [a,b], with a par-

tition, D : a = x

< x

< ··· < x

= b. Let:

( f , D) =

∑

i=1

f (x

) − f (x

i−1

)

(1)

Where V

( f , D) the variation of f (x) with respect

to partition D. If ∃M > 0, to divide everything D,

( f , D) ≤ M. Then, f (x) is called the bounded vari-

ation function on [a, b]. Denote:

( f ) = supV

( f , D) (2)

Let V

( f ) be the total variation of f (x) on [a, b], the

sup here is the upper bound which is the smallest up-

per bound. Based on the total variation, we will deﬁne

pixel ﬂuctuations numbers.

3.1.2 Pixel Fluctuations Numbers

Let f (x) be a function deﬁned on [a, b] when the dis-

tribution is constant:

( f )

max

− P

min

(3)

Where V

is called pixel ﬂuctuations numbers of f (x).

( f ) is the total variation of f (x) on the selected

pixel interval, P

max

is the maximum pixel value of

f (x) on the selected pixel interval, P

min

is the mini-

mum pixel value of f (x) on the selected pixel inter-

val.

In this paper, the pixel interval we choose is the

length of 20 pixels. A pixel interval of a pixel in-

cludes the 10 pixels to the left and right of that pixel.

The pixel ﬂuctuations numbers of that pixel are cal-

culated within this interval. If there are less than 10

pixels to the left or right of a pixel, we supplement

it with the pixel value of that pixel, so that the pixel

interval is also 20 pixels long. Pixel ﬂuctuations num-

bers of each pixel in its pixel interval is calculated line

by line, in order to obtain a pixel ﬂuctuation map of

the same size as the image, which is sent into the gen-

erative network together with the original image.

Pixel ﬂuctuations numbers are calculated which is

the invariant under the afﬁne transformation group,

that is, the translation invariance and the invariance

of the rotation. By adding this feature, more informa-

tion can be provided to the image. We provide a brief

proof:

The afﬁne transformation group contains two ba-

sic transformations:

(1)Scale transformation: T

( f ) = a f (x), a 6= 0

(2)Translation transformation: T

( f ) = f (x) + b

It is easy to know that T 1 ◦ T 1, T 1 ◦ T 2, T 2 ◦

T 1, T 2 ◦ T 2 are still afﬁne transformation. We only

need to prove that the number of ﬂuctuations remains

the same for the two basic transformations.

Let the existence group T

( f ) = a f (x), bring it

into the total variation formula:

( f )) =

( f (x)))



max(T

( f (x)))

− P

min(T

( f (x)))



supV

(a f (x), D)



max(a( f (x)))

− P

min(a( f (x)))



(4)

When a > 0, there are:

( f )) =

asupV

( f (x), D)



max( f (x))

− P

min( f (x))



supV

( f (x), D)

max( f (x))

− P

min( f (x))

= V

( f (x)) (5)

When a < 0, there are:

( f )) =

−asupV

( f (x), D)

−a



max( f (x))

− P

min( f (x))



supV

( f (x), D)

max( f (x))

− P

min( f (x))

= V

( f (x)) (6)

Similarly, when T

( f ) = f (x) + b, it is easy to get:

( f )) = V

( f (x)).

3.2 Generative Adversarial Networks

Figure 2: The training process of the generative adversarial

network.

The Generative Adversarial Network(GAN) was pro-

posed by Goodfellow et al. in 2014. The idea of GAN

is derived from game theory. It proposes the idea of

designing two different networks, one of which is a

generative network to generate a target with a speciﬁc

GAPF: Curve Text Detection based on Generative Adversarial Networks and Pixel Fluctuations

547

meaning image; another network is used as a discrim-

inative network to determine whether the input image

is a network-generated image or a real image. In this

way, the two networks produce adversarial training.

The discriminative network is a binary classiﬁca-

tion network: if the input image is a real image, its

output is 1; if it is a target image generated by the

generation network, its output is 0. During training,

the generative network constantly adjusts its param-

eters to make the generated image as similar to the

real image as possible, which results in the discrimi-

native network not being able to correctly distinguish

whether it is the generated image or the real image.

In contrast, the purpose of the discriminative network

is to distinguish between the two as much as possi-

ble. In this way, after a long period of training, the

GAN ﬁnally reaches the state of Nash equilibrium,

and the results generated by the generative network

can achieve the real effect of falsehood. Assume that

the network has completed the training. The gener-

ated target images (such as segmentation results, col-

oring results, etc.) can be used as correct results.

Its network training process is shown in Figure

2. Differentiable functions G and D in the ﬁgure

represent generative networks and discriminative net-

works, respectively, and z and x represent random

variables and real data, respectively. G(z) refers to

the data generated by the generative network.

3.3 Methodology

3.3.1 Pipeline

The pipeline of the proposed GAPF is illustrated in

Figure 3. In the stage of pre-processing the pictures,

we generate the pixel ﬂuctuations map of each im-

age. Two entries are deﬁned: the original image of

the input image and the pixel ﬂuctuations map of the

original image. At the same time, we also add a pic-

ture entry for discriminating the network to load pixel

ﬂuctuations information.

In the ﬁgure, G represents the generative network

we use to generate the image semantic segmentation

results, x represents the real picture input to the gener-

ator, and f represents the pixel ﬂuctuations map of the

corresponding image. D represents the discriminative

network in the generative adversarial network frame-

work. The input samples of the discriminator include

two types:

(1) A sample is composed of the original image x,

the pixel ﬂuctuations map f of the original image, and

the semantic segmentation result G (x) generated by

the generator, we call it a negative sample pair. This

is shown in the left part of Figure 3.

(2) A sample is composed of the original image x,

the pixel ﬂuctuations map f of the original image, and

the artiﬁcially labeled semantic segmentation result y

of the image, which is called a positive sample pair.

As shown on the right side of Figure 3, we deﬁne the

loss function as follows:

Loss = E

x+ f ,y

[logD(x + f , y)]+E

x+ f ,G(x)

[log

(1 − D(x + f , G(x))] (7)

The training process of D is to maximize the accuracy,

that is, to minimize the loss function. The output of

positive samples tends to 1 and the output of negative

samples tends to 0, so that the overall loss of D tends

to 0. The purpose of G training is to minimize the

accuracy of D, that is, the output of negative samples

approaches 1.

3.3.2 Generative Network

The whole generative network is shown in Figure

4. Inspired by FPN (Lin,2017) and U-net (Ron-

neberger,2015), we adopt a scheme that gradually

merges features from different levels of the stem net-

work. We choose ResNet-50 (He,2016) as our stem

network for the sake of direct and fair comparison

with other methods.

As for the feature merging network, several stages

are stacked sequentially, each consisting of a merging

Figure 3: Illustration of our overall pipeline.

VISAPP 2021 - 16th International Conference on Computer Vision Theory and Applications

548

unit that takes feature maps from the last stage and

corresponding stem network layer. Merging unit is

deﬁned by the following equations:

= f

(8)

= conv

3×3

(conv

1×1

[ f

;U pSampling

×2

i−1

)]),

f or i ≥ 2 (9)

where f

denotes the feature maps of the i-th stage

in the stem network and h

is the feature maps of the

corresponding merging units. In our experiments, up-

sampling is implemented as a deconvolutional layer

as proposed in (Zeiler,2010). After the merging, we

obtain a feature map whose size is 1/2 of the input

images. We apply an additional upsampling layer

and two convolutional layers to produce dense pre-

dictions:

f inal

= U pSampling

×2

) (10)

P = conv

1×1

(conv

3×3

f inal

)) (11)

where P ∈ R

h×w×2

, these two channels represent text

or non-text areas. As a result of the additional upsam-

pling layer, P has the same size as the input image.

The ﬁnal predictions are obtained by taking the soft-

max algorithm.

3.3.3 Discriminative Network

Compared to the generative network, the discrimina-

tive network is simpler. The purpose of the discrimi-

native network is to distinguish between positive and

negative sample pairs, which is a binary classiﬁcation

problem. For the discriminator, we use a simple con-

volutional network structure, as shown in Figure 5:

The discriminator combines two inputs: the real

image and the tag image corresponding to the real im-

age (the generator generates the image or manually

annotated image). Using convolution to continuously

extract high-dimensional features from the input. Fi-

nally, sigmoid is used to classify the results.

3.4 Label Generation

Figure 6 shows the label generation process. Figure

6(a) is an example of manually annotated text, shown

in a red border-box. The sample points in the solid red

box are deﬁned as positive samples with a value of 1,

while the other sample points are deﬁned as negative

samples with a value of 0. The ﬁnal result is shown in

Figure 6(b).

(a) (b)

Figure 6: The illustration of label generation. (a) shows the

original text instances. (b) shows the segmentation masks.

Figure 4: The architecture of generative Network.

Figure 5: The architecture of discriminative Network.

GAPF: Curve Text Detection based on Generative Adversarial Networks and Pixel Fluctuations

549

3.5 Text Boundary Generation

After adversarial learning of the generative network

and the discriminative network, the generative net-

work generates accurate segmentation results. And

our goal is to get the bounding box that surrounds

the text area. The speciﬁc implementation details

are shown in Algorithm 1. We chose n positive pix-

els in each text region and use the principal curve

(Trevor,1989) to regress the curve centerline. Seven

points are chosen from the centerline. For each pair

of points that are adjacent in the centerline, we use the

center point of two points as a rectangle center and

generate a circumscribed rectangle of the area where

are the text region between the pare center points. We

will get the boundary of the text region by repeating

these steps.

Algorithm 1: Text Boundary Generation.

Input: Segmentation result set S for a given image

Output: Text boundary set B for the image

1: B = {}

2: for each segmentation result s ∈ S do

3: Boundary b = [ ]

4: choose n positive pixels

5: use principal curve to ﬁnd curve center line

6: choose 7 point p

(i = 0 · ·· 6) to represent center

line

7: for i ∈ [1, 6] do

8: generate circumscribed rectangle of the area be-

tween b

and b

i+1

9: regard rectangle points as boundary points

10: insert the left two points into b

11: end for

12: insert the b into B

13: end for

14: return B

4 EXPERIMENT

4.1 Datasets

CTW1500 (Liu,2017) consists of 1000 training and

500 testing images. Each text instance annotation is a

polygon with 14 vertexes to deﬁne the text region at

the level of the text line. The text instances include

both inclined texts as well as the horizontal texts.

Total-Text(Chee,2017) is a newly-released dataset

for curve text detection. Horizontal, multi-Oriented

and curve text instances are contained in Total-Text.

The benchmark consists of 1255 training images and

300 testing images. The images are annotated at the

level of the word by a polygon with 2N vertices (N

∈ 2, ...,15).

MSRA-TD500 (Karatzas,2015) contains 500 nat-

ural images, which are split into 300 training images

and 200 testing images, collected both indoors and

outdoors using a pocket camera. The images contain

English and Chinese scripts. Text regions are anno-

tated by rotated rectangles.

ICDAR2015 (Yao,2012) was introduced in the IC-

DAR 2015 Robust Reading Competition for inciden-

tal scene text detection, consisting of 1000 training

images and 500 testing images, both with texts in En-

glish. The annotations are at the word level using

quadrilateral boxes.

4.2 Implementation Details

The generator and discriminator perform adversarial

training. First, the generator is ﬁxed to train the dis-

criminator, then the discriminator is ﬁxed to train the

generator, and then the loop training is continued.

Through adversarial learning, the capabilities of the

generator’s discriminator have been enhanced. In the

end, the discriminator cannot distinguish between the

segmentation result generated by the generator and

the real segmentation result. At this point, the training

is over. When a new image is input, the segmentation

result generated by the generator can be used as the

correct text segmentation result.

Our method is implemented in Pytorch1.1.

Speciﬁcally, our network is trained with stochastic

gradient descent (SGD) for 100K iterations with the

initial learning rate being 0.0001 and a minibatch of

6 images. Weight decay and momentum are set as

0.0005 and 0.9. In terms of initial assignment, gaus-

sian distribution with a mean of 0 and a variance of

0.01 is used for random initialization.

4.3 Results and Comparison

To test the ability of curve text detection, we evalu-

ate our method on CTW1500 and Total-Text, which

mainly contains the curve texts.

Table 1: Quantitative results of different methods evaluated

on CTW1500.

Method Precision Recall F-measure

CTD (Liu,2017) 74.3 65.2 69.5

CTD+TLOC (Liu,2017) 77.4 69.8 73.4

SLPR (Zhu,2018) 80.1 70.1 74.8

TextSnake (Long,2018) 67.9 85.3 75.6

PSENet (Li,2019) 82.09 77.84 79.9

LSAE (Tian,2019) 82.7 77.8 80.1

GAPF 83.2 79.6 81.3

VISAPP 2021 - 16th International Conference on Computer Vision Theory and Applications

550

Table 2: Quantitative results of different methods evaluated

on Total-Text.

Method Precision Recall F-measure

MaskSpotter (Lyu,2018) 69.0 55.0 61.3

TextSnake (Long,2018) 82.7 74.5 78.4

PSENet (Li,2019) 84.54 75.23 79.61

GAPF 83.6 76.8 80.05

Besides, to prove the universality of the proposed

GAPF in this paper, we evaluate the proposed method

on the ICDAR2015 and MSRA-TD500 to test its abil-

ity for oriented text detection. Table 3 and Table 4

show the comparison results with advanced methods.

The results show that the method in this paper has

an accurate detection effect on the horizontal and in-

clined text.

Table 3: Quantitative results of different methods evaluated

on Total-Text.

Method Precision Recall F-measure

CCNF (Yao,2016) 72.86 58.69 64.77

RRPN (Ma.2018) 82.17 73.23 77.44

EAST (Zhou,2017) 83.27 78.33 80.72

TextSnake (Long,2018) 84.90 80.40 82.60

PixelLink (Dan,2018) 85.50 82.00 83.70

GAPF 86.42 81.83 84.06

Table 4: Quantitative results of different methods evaluated

on MSRA-TD500.

Method Precision Recall F-measure

RRPN (Ma.2018) 82 69 75

EAST (Zhou,2017) 83.56 67.13 74.45

TextSnake (Long,2018) 83.2 73.9 78.3

PixelLink (Dan,2018) 83 73.2 77.8

GAPF 84.2 73.6 78.54

Figure 7 depicts several detection examples by the

proposed GAPF. The solid red line box in the ﬁgure

represents the output text detection box. It can be seen

that the detection results of the method GAPF are bet-

ter on these curved texts and can also effectively sup-

port horizontal and multi-direction text detection.

Figure 7: Qualitative results by the proposed method. From

top to bottom in row: image from CTW1500, Total-Text,

ICDAR 2015, and MSRA-TD500.

5 CONCLUSION

We propose a text detection framework (GAPF) for

the arbitrary shape scene text. The text segmentation

map is generated by fusing pixel ﬂuctuation informa-

tion into the generation adversarial network. Then use

the boundary generation algorithm to get the bound-

ing box of the text area. Our method is robust to

shapes and can easily detect text instances of arbitrary

shapes. Experimental comparisons with the state-of-

the-art approaches on multiple datasets show the ef-

fectiveness of the proposed GAPF for the text detec-

tion task.

REFERENCES

Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian

Sun. Deep residual learning for image recognition. In

CVPR, 770–778(2016)

Goodfellow I J, Pougetabadie J, Mirza M, et al. Generative

Adversarial Networks[J]. Advances in Neural Infor-

mation Processing Systems, 2672–2680(2014)

GAPF: Curve Text Detection based on Generative Adversarial Networks and Pixel Fluctuations

551

Ren, S., He, K., Girshick, R., Sun, J.: Faster R-CNN: To-

wards Real-Time Object Detection with Region Pro-

posal Networks. IEEE Trans. Pattern Anal. Mach. In-

tell. 39(6), 1137–1149 (2017)

Ma, J., Shao, W., Ye, H., Wang, L., Wang, H., Zheng, Y.,

Xue, X.: Arbitrary-Oriented Scene Text Detection via

Rotation Proposals. IEEE Trans. Multimedia 20(11),

3111–3122 (2018)

Liao, M., Shi, B., Bai, X., Wang, X., Liu, W.: TextBoxes:

A Fast Text Detector with a Single Deep Neural Net-

work. In Proc. AAAI, 4161–4167(2017)

Liu, W., Anguelov, D., Erhan, D., Szegedy, C., Reed, S., Fu,

C., Berg, A.: SSD: Single Shot MultiBox Detector. In

Proc.ECCV, 21–37(2016)

Zhou, X., Yao, C., Wen, H., Wang, Y., Zhou, S., He, W.,

Liang, J.: EAST: An Efﬁcient and Accurate Scene

Text Detector. In Proc.CVPR, 2642–2651(2017)

Zhi Tian, Weilin Huang, Tong He, Pan He, and Yu Qiao,

”Detecting text in natural image with connectionist

text proposal network,” in ECCV, 56–72(2016)

Alex Graves and J

urgen Schmidhuber, ”Framewise

phoneme classiﬁcation with bidirectional lstm and

other neural network architectures”, Neural Networks,

vol.18, no.5-6, pp.602–610 (2005)

Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully

convolutional networks for semantic segmentation. In

CVPR, 3431–3440(2015)

Zheng Zhang, Chengquan Zhang, Wei Shen, Cong Yao,

Wenyu Liu, and Xiang Bai. Multi-oriented text de-

tection with fully convolutional networks. In CVPR,

4159–4167(2016)

Dan Deng, Haifeng Liu, Xuelong Li, and Deng Cai. Pix-

ellink: Detecting scene text via instance segmentation.

In AAAI, 6773–6780(2018)

Shangbang Long, Jiaqiang Ruan, Wenjie Zhang, Xin He,

Wenhao Wu, and Cong Yao. Textsnake: A ﬂexible

representation for detecting text of arbitrary shapes.

In ECCV,19–35(2018)

Xiang Li, Wenhai Wang, Wenbo Hou, Ruo-Ze Liu, Tong

Lu, and Jian Yang, ”Shape robust text detection

with progressive scale expansion network,” In CVPR,

9336–9345(2019).

Lin, T.Y., Dollar, P., Girshick, R., He, K., Hariharan, B.,

Belongie, S.: Feature pyramid networks for object de-

tection. In: The IEEE Conference on Computer Vision

and Pattern Recognition. In CVPR, 936–944(2017)

Ronneberger, O., Fischer, P., Brox, T.: U-Net: Convolu-

tional Networks for Biomedical Image Segmentation.

Springer International Publishing. In MICCAI, 234–

24 (2015)

He, K., Zhang, X., Ren, S., Sun, J.: Deep residual learning

for image recognition. In CVPR, 770–778(2016)

Zeiler, M.D., Krishnan, D., Taylor, G.W., Fergus, R.: De-

convolutional networks. In CVPR, 2528–2535 (2010)

Trevor Hastie and Werner Stuetzle, ”Principal curves”,

Journal of the American Statistical Association, vol.

84, no. 406, pp. 502–516(1989)

Yuliang Liu, Lianwen Jin, Shuaitao Zhang, and Sheng

Zhang, ”Detecting curve text in the wild: New dataset

and new solution”, arXiv:1712.02170, (2017)

Chee Kheng Ch’ng and Chee Seng Chan, ”Total-text: A

comprehensive dataset for scene text detection and

recognition”, in ICDAR, 935–942 (2017)

Karatzas, D., Gomez, L., Nicolaou, A., Ghosh, S., Bag-

danov, A., Iwamura, M., Matas, J., Neumann, L.,

Chandrasekhar, V., Lu, S., Shafait, F., Uchida, S., Val-

veny, E.: ICDAR 2015 competition on robust reading.

In Proc.ICDAR, 1156–1160(2015)

Yao, C., Bai, X., Liu, W., Ma, Y., Tu, Z.: Detecting texts

of arbitrary orientations in natural images. In CVPR,

1083–1090(2012)

Yuliang Liu, Lianwen Jin, Shuaitao Zhang, and Sheng

Zhang, ”Detecting curve text in the wild: New dataset

and new solution”, arXiv:1712.02170, (2017)

Yixing Zhu and Jun Du, ”Sliding line point regression for

shape robust scene text detection”, arXiv:1801.09969,

(2018)

Zhuotao Tian, Michelle Shu, Pengyuan Lyu, Ruiyu Li,

Chao Zhou, Xiaoyong Shen, Jiaya Jia:Learning

Shape-Aware Embedding for Scene Text Detection. In

CVPR, 4234–4243(2019)

Pengyuan Lyu, Minghui Liao, Cong Yao, Wenhao Wu, and

Xiang Bai, ”Mask textspotter: An end-to-end train-

able neural network for spotting text with arbitrary

shapes”,in ECCV,71–88(2018)

Yao, C., Bai, X., Sang, N., Zhou, X., Zhou, S., Cao, Z.:

Scene Text Detection via Holistic, Multi-Channel Pre-

diction. CoRR abs/1606.09002 (2016)

VISAPP 2021 - 16th International Conference on Computer Vision Theory and Applications

552