txt/2201.00132.txt

SAFL: A Self-Attention Scene Text Recognizer
with Focal Loss
Bao Hieu Tran, Thanh Le-Cong, Huu Manh Nguyen, Duc Anh Le, Thanh Hung Nguyeny, Phi Le Nguyeny
School of Information and Communication Technology
Hanoi University of Science and Technology
Hanoi, Vietnam
fhieu.tb167182, thanh.ld164834, manh.nh166428, anh.nd160126 [email protected],flenp, [email protected]
Abstract —In the last decades, scene text recognition has gained
worldwide attention from both the academic community and
actual users due to its importance in a wide range of applications.
Despite achievements in optical character recognition, scene
text recognition remains challenging due to inherent problems
such as distortions or irregular layout. Most of the existing
approaches mainly leverage recurrence or convolution-based
neural networks. However, while recurrent neural networks
(RNNs) usually suffer from slow training speed due to sequential
computation and encounter problems as vanishing gradient or
bottleneck, CNN endures a trade-off between complexity and
performance. In this paper, we introduce SAFL, a self-attention-
based neural network model with the focal loss for scene text
recognition, to overcome the limitation of the existing approaches.
The use of focal loss instead of negative log-likelihood helps the
model focus more on low-frequency samples training. Moreover,
to deal with the distortions and irregular texts, we exploit Spatial
TransformerNetwork (STN) to rectify text before passing to the
recognition network. We perform experiments to compare the
performance of the proposed model with seven benchmarks.
The numerical results show that our model achieves the best
performance.
Index Terms —Scene Text Recognition, Self-attention, Focal
loss,
I. I NTRODUCTION
In recent years, text recognition has attracted the attention
of both academia and actual users due to its application
on various domains such as translation in mixed reality,
autonomous driving, or assistive technology for the blind.
Text recognition can be classified into two main categories:
scanned document recognition and scene text recognition.
While the former has achieved significant advancements, the
latter remains challenging due to scene texts’ inherent char-
acteristics such as the distortion and irregular shapes of the
texts. Recent methods in scene text recognition are inspired
by the success of deep learning-based recognition models.
Generally, these methods can be classified in two approaches:
recurrent neural networks (RNN) based and convolutional
neural networks (CNN) based. RNN-based models have shown
their effectiveness, thanks to capturing contextual information
and dependencies between different patches. However, RNNs
typically compute along with the symbol positions of the
input and output sequences, which cannot be performed in
Authors contribute equallyyCorresponding authorparallel fashion, thus leads to high training time. Furthermore,
RNNs also encounter problems such as vanishing gradient
[1] or bottleneck [2]. CNN-based approach, which allows
computing the hidden representation parallelly, have been
proposed to speed up the training procedure. However, to
capture the dependencies between distant patches in long input
sequences, CNN models require stacking more convolutional
layers, which significantly increases the network’s complexity.
Therefore, CNN-based methods suffer the trade-off between
complexity and accuracy. To remedy these limitations, in nat-
ural language processing (NLP) fields, a self-attention based
mechanism named transformer [3] has been proposed. In the
transformer, dependencies between different input and output
positions are captured using a self-attention mechanism instead
of sequential procedures in RNN. This mechanism allows
more computation parallelization with higher performance. In
the computer vision domain, some research have leveraged the
transformer architecture and showed the effectiveness of some
problems [4] [5]
Inspired by the transformer network, in this paper, we
propose a self-attention based scene text recognizer with focal
loss, namely as SAFL. Moreover, to tackle irregular shapes of
scene texts, we also exploit a text rectification named Spatial
Transformer Network (STN) to enhance the quality of text
before passing to the recognition network. SAFL, as depicted
in Figure 1, contains three components: rectification, feature
extraction, and recognition. First, given an input image, the
rectification network, built based on the Spatial Transformer
Network (STN) [6], transforms the image to rectify its text.
Then, the features of the rectified image are extracted using
a convolutional neural network. Finally, a self-attention based
recognition network is applied to predict the output character
sequence. Specifically, the recognition network is an encoder-
decoder model, where the encoder utilizes multi-head self-
attention to transform input sequence to hidden feature rep-
resentation, then the decoder applies another multi-head self-
attention to output character sequence. To balance the training
data for improving the prediction accuracy, we exploit focal
loss instead of negative log-likelihood as in most recent works
[7] [8].
To evaluate our proposed model’s performance, we train
SAFL with two synthetic datasets: Synth90k [9] and SynthText
[10], and compare its accuracy with standard benchmarks,arXiv:2201.00132v1  [cs.CV]  1 Jan 2022Fig. 1. Overview of SAFL
on both regular and irregular datasets. The experiment results
show that our method outperforms the state-of-the-art on all
datasets. Furthermore, we also perform experiments to study
the effectiveness of focal loss. The numerical results show the
superiority of focal loss over the negative log-likelihood loss
on all datasets.
The remainder of the paper is organized as follows. Section
II introduces related works. We describe the details of the
proposed model in Section III and present the evaluation
results in Section IV. Finally, we conclude the paper and
discuss the future works in Section V.
II. R ELATED WORK
Scene text recognition has attracted great interest over
the past few years. Comprehensive surveys for scene text
recognition may be found in [11] [12] [13]. As categorized
by previous works [8] [14] [15], scene text may be divided
into two categories: regular and irregular text. The regular text
usually has a nearly horizontal shape, while the irregular text
has an arbitrary shape, which may be distorted.
A. Regular text recognition
Early work mainly focused on regular text and used a
bottom-up scheme, which first detects individual characters
using a sliding window, then recognizing the characters us-
ing dynamic programming or lexicon search [16] [17] [18].
However, these methods have an inherent limitation, which is
ignoring contextual dependencies between characters. Shi et al.
[19] and He et al. [20] typically regard text recognition as a
sequence-to-sequence problem. Input images and output texts
are typically represented as patch sequences and character
sequences, respectively. This technique allows leveraging deep
learning techniques such as RNNs or CNNs to capture con-
textual dependencies between characters [7] [19] [20], lead to
significant improvements in accuracy on standard benchmarks.
Therefore, recent work has shifted focus to the irregular text,
a more challenging problem of scene text recognition.
B. Irregular text recognition
Irregular text is a recent challenging problem of scene text
recognition, which refers to texts with perspective distortionsand arbitrary shape. The early works correct perspective distor-
tions by using hand-craft features. However, these approaches
require correct tunning by expert knowledge for achieving
the best results because of a large variety of hyperparame-
ters. Recently, Yang et al. [21] proposed an auxiliary dense
character detection model and an alignment loss to effectively
solve irregular text problems. Liu et al. [22] introduced a
Character-Aware Neural Network (Char-Net) to detect and rec-
tify individual characters. Shi et al. [7] [8] addressed irregular
text problems with a rectification network based on Spatial
Transformer Network (STN), which transform input image for
better recognition. Zhan et al. [23] proposed a rectification
network employing a novel line-fitting transformation and an
iterative rectification pipeline for correction of perspective and
curvature distortions of irregular texts.
III. P ROPOSED MODEL
Figure 1 shows the structure of SAFL, which is comprised
of three main components: rectification, feature extraction, and
recognition. The rectification module is a Spatial Transformer
Network (STN) [6], which receives the original image and
rectifies the text to enhance the quality. The feature extraction
module is a convolution neural network that extracts the
information of the rectified image and represents it into a
vector sequence. The final module, i.e., recognition, is based
on the self-attention mechanism and the transformer network
architecture [3], to predict character sequence from the feature
sequence. In the following, we first present the details of the
three components in Section III-A, III-B and III-C, respec-
tively. Then, we describe the training strategy using focal loss
in Section III-D.
A. Rectification
In this module, we leverage a Thin Plate Spline (TPS) trans-
formation [8], a variant of STN, to construct a rectification
network. Given the input image Iwith an arbitrary size, the
rectification module first resizes Iinto a predefined fixed size.
Then the module detects several control points along the top
and bottom of the text’s bounding. Finally, TPS applies asmooth spline interpolation between a set of control points
to rectify the predicted region to obtain a fixed-size image.
B. Feature Extraction
We exploit the convolution neural network (CNN) to extract
the features of the rectified image (obtained from a rectification
network) into a sequence of vectors. Specifically, the input
image is passed through convolution layers (ConvNet) to
produce a feature map. Then, the model separates the feature
map by rows. The output received after separating the feature
map are feature vectors arranged in sequences. The scene
text recognition problem then becomes a sequence-to-sequence
problem whose input is a sequence of characteristic vectors,
and whose output is a sequence of characters predicted. Based
on the proposal in [3], we further improve information about
the position of the text in the input image by using positional
encoding. Each position posis represented by a vector whose
value of the ithdimension, i.e., PE (pos;i ), is defined as
PE (pos;i )=8
<
:sinpos
100002i
dmodel;if0idmodel
2
cospos
100002i
dmodel;ifdmodel
2idmodel;
(1)
wheredmodel is the vector size. The position information is
added into the encoding vectors.
C. Self-attention based recognition network
The architecture of the recognition network follows the
encoder-decoder model. Both encoder blocks and decoder
blocks are built based on the self-attention mechanism. We
will briefly review this mechanism before describing each
network’s details.
1) Self-attention mechanism: Self-attention is a mechanism
that extracts the correlation between different positions of a
single sequence to compute a representation of the sequence.
In this paper, we utilize the scaled dot-product attention
proposed in [3]. This mechanism consists of queries and keys
of dimension dk, and values of dimension dv. Each query
performs the dot product of all keys to obtain their correlation.
Then, we obtain the weights on the values by using the softmax
function. In practice, the keys, values, and queries are also
packed together into matrices K,VandQ. The matrix of the
outputs is computed as follow:
Attention (Q;K;V ) =softmaxQKT
pdk
V (2)
The dot product is scaled by1pdkto alleviate the small softmax
values which lead to extremely small gradients with large
values ofdk. [3].
2) Encoder: Encoder is a stack of Neblocks. Each block
consists of two main layers. The first layer is a multi-head
attention layer, and the second layer is a fully-connected feed-
forward layer. The multi-head attention layer is the combi-
nation of multiple outputs of the scale dot product attention.
Each scale-dot product attention returns a matrix representing
the feature sequences, which is called head attention. The
combination of multiple head attentions to the multi-head
Fig. 2. Frenquency of characters in training lexicon
attention allows our model to learn more representations of
feature sequences, thereby increasing the diversity of the
extracted information, and thereby enhance the performance.
Multi-head attention can be formulated as follows:
MultiHead (Q;K;V ) =Concat (head 1;:::;head h)WO
(3)
whereheadi=Attention
QWQ
i;KWK
i;VWV
i
,his the
number of heads, WQ
i2Rdmodeldk;WK
i2Rdmodeldk;WV
i2
Rdmodeldv,WO2Rhdvdmodel are weight matrices. dk;dv
anddmodel are set to the same value. Layer normalization
[24] and residual connection [25] are added into each main
layer (i.e., multi-head attention layer and fully-connected
layer) to improve the training effect. Specifically, the residual
connections helps to decrease the loss of information in the
backpropagation process, while the normalization makes the
training process more stable. Consequently, the output of
each main layer with the input xcan be represented as
LayerNorm (x+Layer (x)), whereLayer (x)is the function
implemented by the layer itself, and LayerNorm ()represents
the normalization operation. The blocks of the encoder are
stacked sequentially, i.e., the output of the previous block is
the input of the following block.
3) Decoder: The decoding process predicts the words in
a sentence from left to right, starting with the hstartitag
until encountering the henditag. The decoder is comprised of
Nddecoder blocks. Each block is also built based on multi-
head attention and a fully connected layer. The multi-head
attention in the decoder does not consider words that have
not been predicted by weighting these positions with 1.
Furthermore, the decoder uses additional multi-head attention
that receives keys and values from the encoder and queries
from the decoder. Finally, the decoder’s output is converted
into a probability distribution through a linear transformation
and softmax function.
D. Training
Figure 2 shows that the lexicon of training datasets suffers
from an unbalanced sample distribution. The unbalance may
lead to severe overfitting for high-frequency samples and un-
derfitting for low-frequency samples. To this end, we propose
to use focal loss [26] instead of negative log-likelihood as in
most of recent methods [7] [8]. By exploiting focal loss, themodel will not encounter the phenomenon of ignoring to train
low-frequency samples.
Focal loss is known as an effective loss function to address
the unbalance of datasets. By reshaping the standard cross-
entropy loss, focal loss reduces the impacts of high-frequency
samples and thus focus training on low-frequency ones [26].
The focal loss is defined as follows:
FL(pt) =t(1pt)
log (pt); (4)
where,ptis the probability of the predicted value, computed
using softmax function, and
are tunable hyperparameters
used to balance the loss. Intuitively, focal loss is obtained
by multiplying cross entropy by t(1pt)
. Note that the
weightt(1pt)
is inversely proportional with pt, thus
the focal loss helps to reduce the impact of high-frequency
samples (whose value of ptis usually high) and focus more
on low-frequency ones (which usually have low value of pt).
Based on focal loss, we define our training objective as
follows:
L=TX
t=1(t(1p(ytjI))
logp(ytjI)))) (5)
whereytare the predicted characters, Tis the length of the
predicted sequence, and Iis the input image.
IV. P ERFORMANCE EVALUATION
In this section, we conduct experiments to demonstrate the
effectiveness of our proposed model. We first briefly introduce
datasets used for training and testing, then we describe our
implementation details. Next, we analyze the effect of focal
loss on our model. Finally, we compare our model against
state-of-the-art techniques on seven public benchmark datasets,
including regular and irregular text.
A. Datasets
The training datasets contains two datasets: Synth90k and
SynthText . Synth90k is a synthetic dataset introduced in [9].
This dataset contains 9 million images created by combining
90.000 common English words and random variations and
effects. SynthText is a synthetic dataset introduced in [10],
which contains 7 million samples by the same generation
process as Synth90k [9]. However, SynthText is targeted for
text detection so that an image may contain several words. All
experiments are evaluated on seven well-known public bench-
marks described, which can be divided into two categories:
regular text and irregular text. Regular text datasets include
IIIT5K, SVT, ICDAR03, ICDAR13.
IIIT5K [27] contains 3000 test images collected from
Google image searches.
ICDAR03 [28] contains 860 word-box cropped images.
ICDAR13 [29] contains 1015 word-box cropped images.
SVT contains 647 testing word-box collected from
Google Street View.
Irregular text datasets include ICDAR15, SVT-P, CUTE.ICDAR15 [30] contains 1811 testing word-box cropped
images collected from Google Glass without careful
positioning and focusing.
SVT-P [31] contains 645 testing word-box cropped im-
ages collected from Google Street View. Most of them
are heavily distorted by the non-frontal view angle.
CUTE [32] contains 288 word-box cropped images,
which are curved text images.
B. Configurations
1) Implementation Detail: We implement the proposed
model by Pytorch library and Python programming language.
The model is trained and tested on an NDIVIA RTX 2080 Ti
GPU with 12GB memory. We train the model from scratch
using Adam optimizer with the learning rate of 0:00002 .
To evaluate the trained model, we use dataset III5K. The
pretrained model and code are available at [33]
2) Rectification Network: All input images are resized
to64256 before applying the rectification network. The
rectification network consists of three components: a localiza-
tion network, a thin plate spline (TPS) transformation, and a
sampler. The localization network consists of 6 convolutional
layers with the kernel size of 33and two fully-connected
(FCN) layers. Each FCN is followed by a 22max-pooling
layer. The number of the output filters is 32, 64, 128, 256,
and 256. The number of output units of FCN is 512 and 2K,
respectively, where Kis the number of the control points. In
all experiments, we set Kto 20, as suggested by [8]. The
sampler generates the rectified image with a size of 32100.
The size of the rectified image is also the input size of the
feature extraction module.
3) Feature Extraction: We construct the feature extraction
module based on Resnet architecture [25]. The configurations
of the feature extraction network are listed in Table I. Our
feature extraction network contains five blocks of 45 residual
layers. Each residual unit consists of a 11convolutional
layer, followed by a 33convolution layer. In the first
two blocks, we use 22stride to reduce the feature map
dimension. In the next blocks, we use 21stride to down-
sampled feature maps. The 21stride also allows us to
retain more information horizontally to distinguish neighbor
characters effectively.
TABLE I
FEATURE EXTRACTION NETWORK CONFIGURATIONS .EACH BLOCK IS A
RESIDUAL NETWORK BLOCK . ”S”STANDS FOR STRIDE OF THE FIRST
CONVOLUTIONAL LAYER IN A BLOCK .
Layer Feature map size Configuration
EncoderBlock 0 32100 33conv, s( 11)
Block 1 1650
11 conv ;32
33 conv ;32
3, s(22)
Block 2 825
11 conv ;64
33 conv ;32
3, s(22)
Block 3 425
11 conv ;128
33 conv ;32
3, s(21)
Block 4 225
11 conv ;256
33 conv ;32
3, s(21)
Block 5 125
11 conv ;512
33 conv ;32
3, s(21)4) Recognition: The number of blocks in the encoder and
the decoder are set both to 4. In each block of the encoder and
the decoder, the dimension of the feed forward vector and the
ouput vector are set to 2048 and512, respectively. The number
of head attention layers is set to 8. The decoder recognizes 94
different characters, including numbers, alphabet characters,
uppercase, lowercase, and 32 punctuation in ASCII.
C. Result and Discussion
1) Impact of focal loss: To analyze the effect of focal loss,
we study two variants of the proposed model. The first variant
uses negative log-likelihood, and the second one leverages
focal loss.
TABLE II
RECOGNITION ACCURACIES WITH NEGATIVE LOG -LIKELIHOOD AND
FOCAL LOSS
Variant Negative log-likelihood Focal Loss
IIIT5K 92.6 93.9
SVT 85.8 88.6
ICDAR03 94.1 95
ICDAR13 92 92.8
ICDAR15 76.1 77.5
SVT-P 79.4 81.7
CUTE 80.6 85.4
Avarage 86.9 88.2
As shown in Table II, the model with focal loss outperforms
the one with log-likelihood on all datasets. Notably, on aver-
age, focal loss improves the accuracy by 2.3 % compared to
log-likelihood. For the best case, i.e., CUTE, the performance
gap between the two variants is 4.8 %
2) Impact of rectification network: In this section, we study
the effect of text rectification by comparing SAFL and a
variant which does not include the rectification module.
TABLE III
RECOGNITION ACCURACIES WITH AND WITHOUT RECTIFICATION
Variant SAFL w/o text rectification SAFL
IIIT5K 90.7 93.9
SVT 83.3 88.6
ICDAR03 93 95
ICDAR13 90.7 92.8
ICDAR15 72.9 77.5
SVT-P 71.6 81.7
CUTE 77.4 85.4
Avarage 84.1 88.2
Table III depicts the recognition accuracies of the two mod-
els over seven datasets. It can be observed that the rectification
module increases the accuracy significantly. Specifically, the
performance gap between SAFL and the one without the
rectification module is 4:1%on average. In the best cases,
SAFL improves the accuracy by 10:1%and7%compared
to the other on the datasets SVT-P and CUTE, respectively.
The reason is that both SVT-P and CUTE contains many both
irregular texts such as perspective texts or curved texts.
3) Comparison with State-of-the-art: In this section, we
compare the performance of SAFL with the latest approaches
in scene text recognition. The evaluation results are shownin Table IV. In each column, the best value is bolded.
the ”Avarage” column is the weighted average over all the
data sets. Concerning the irregular text, it can be observed
that SAFL achieves the best performance on 3data sets.
Particularly, SAFL outperforms the current state-of-the-art,
ESIR [23], by a margin of 1:2% on average, particulary
on CUTE ( +2:1%) and SVT-P ( +2:1%). Concerning the
regular datasets, SAFL outperforms the other methods on two
datasets IIIT5K and ICDAR03. Moreover, SAFL also shows
the highest average accuracy over all the regular text datasets.
To summarize, SAFL achieves the best performance on 5 of
7 datasets and the highest average accuracy on both irregular
and regular texts.
V. C ONCLUSION
In this paper, we proposed SAFL, a deep learning model
for scene text recognition, which exploits self-attention mecha-
nism and focal loss. The experiment results showed that SAFL
achieves the highest average accuracy on both the regular
datasets and irregular datasets. Moreover, SAFL outperforms
the state-of-the-art on CUTE dataset by a margin of 2:1%.
Summary, SAFL shows superior performance on 5 out of 7
benchmarks, including IIIT5k, ICDAR 2003, ICDAR 2015,
SVT-P and CUTE.
ACKNOWLEDGMENT
We would like to thank AIMENEXT Co., Ltd. for support-
ing our research.
REFERENCES
[1] Y . Bengio, P. Simard, and P. Frasconi, “Learning long-term dependencies
with gradient descent is difficult,” IEEE transactions on neural networks ,
vol. 5, no. 2, pp. 157–166, 1994.
[2] K. Cho, B. Van Merri ¨enboer, D. Bahdanau, and Y . Bengio, “On the
properties of neural machine translation: Encoder-decoder approaches,”
arXiv preprint arXiv:1409.1259 , 2014.
[3] A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez,
Ł. Kaiser, and I. Polosukhin, “Attention is all you need,” in Advances
in neural information processing systems , 2017, pp. 5998–6008.
[4] N. Carion, F. Massa, G. Synnaeve, N. Usunier, A. Kirillov, and
S. Zagoruyko, “End-to-end object detection with transformers,” arXiv
preprint arXiv:2005.12872 , 2020.
[5] N. Parmar, A. Vaswani, J. Uszkoreit, Ł. Kaiser, N. Shazeer, A. Ku, and
D. Tran, “Image transformer,” arXiv preprint arXiv:1802.05751 , 2018.
[6] M. Jaderberg, K. Simonyan, A. Zisserman et al. , “Spatial transformer
networks,” in Advances in neural information processing systems , 2015,
pp. 2017–2025.
[7] B. Shi, X. Wang, P. Lyu, C. Yao, and X. Bai, “Robust scene text
recognition with automatic rectification,” in Proceedings of the IEEE
conference on computer vision and pattern recognition , 2016, pp. 4168–
4176.
[8] B. Shi, M. Yang, X. Wang, P. Lyu, C. Yao, and X. Bai, “Aster:
An attentional scene text recognizer with flexible rectification,” IEEE
transactions on pattern analysis and machine intelligence , vol. 41, no. 9,
pp. 2035–2048, 2018.
[9] M. Jaderberg, K. Simonyan, A. Vedaldi, and A. Zisserman, “Synthetic
data and artificial neural networks for natural scene text recognition,”
arXiv preprint arXiv:1406.2227 , 2014.
[10] A. Gupta, A. Vedaldi, and A. Zisserman, “Synthetic data for text
localisation in natural images,” in Proceedings of the IEEE conference
on computer vision and pattern recognition , 2016, pp. 2315–2324.
[11] K. Simonyan and A. Zisserman, “Very deep convolutional networks for
large-scale image recognition,” arXiv preprint arXiv:1409.1556 , 2014.TABLE IV
SCENE TEXT ACCURANCIES (%) OVER SEVEN PUBLIC BENCHMARK TEST DATASETS .
MethodRegular test dataset Irregular test dataset
IIIT5k SVT ICDAR03 ICDAR13 Average ICDAR15 SVT-P CUTE Average
Jaderberg et al. [34] - 80.7 93.1 90.8 - - - - -
CRNN [19] 78.2 80.8 89.4 86.7 81.8 - - - -
RARE [7] 81.9 81.9 90.1 88.6 85.3 - 71.8 59.2 -
Lee et al. 78.4 80.7 88.7 90.0 82.4 - - - -
Yang et al. [21] - 75.8 - - - - 75.8 69.3 -
FAN [35] 87.4 85.9 94.2 93.3 89.4 70.6 - - -
Shi et al. [7] 81.2 82.7 91.9 89.6 84.6 - - - -
Yang et al. [21] - - - - - - 75.8 69.3 -
Char-Net [22] 83.6 84.4 91.5 90.8 86.2 60.0 73.5 - -
AON [36] 87.0 82.8 91.5 - - 68.2 73.0 76.8 70.0
EP [37] 88.3 87.5 94.6 94.4 90.3 73.9 - - -
Liao et al. [38] 91.9 86.4 - 86.4 - - - 79.9 -
Baek et al. [14] 87.9 87.5 94.9 92.3 89.8 71.8 79.2 74.0 73.6
ASTER [8] 93.4 89.5 94.5 91.8 92.8 76.1 78.5 79.5 76.9
SAR [39] 91.5 84.5 - 91.0 - 69.2 76.4 83.3 72.1
ESIR [23] 93.3 90.2 - 91.3 - 76.9 79.6 83.3 78.1
SAFL 93.9 88.6 95 92.8 93.3 77.5 81.7 85.4 79.3
[12] Y . Zhu, C. Yao, and X. Bai, “Scene text detection and recognition:
Recent advances and future trends,” Frontiers of Computer Science ,
vol. 10, no. 1, pp. 19–36, 2016.
[13] Q. Ye and D. Doermann, “Text detection and recognition in imagery: A
survey,” IEEE transactions on pattern analysis and machine intelligence ,
vol. 37, no. 7, pp. 1480–1500, 2014.
[14] J. Baek, G. Kim, J. Lee, S. Park, D. Han, S. Yun, S. J. Oh, and
H. Lee, “What is wrong with scene text recognition model comparisons?
dataset and model analysis,” in Proceedings of the IEEE International
Conference on Computer Vision , 2019, pp. 4715–4723.
[15] P. Wang, L. Yang, H. Li, Y . Deng, C. Shen, and Y . Zhang, “A simple and
robust convolutional-attention network for irregular text recognition,”
arXiv preprint arXiv:1904.01375 , vol. 6, 2019.
[16] K. Wang, B. Babenko, and S. Belongie, “End-to-end scene text recog-
nition,” in 2011 International Conference on Computer Vision . IEEE,
2011, pp. 1457–1464.
[17] C. Yao, X. Bai, B. Shi, and W. Liu, “Strokelets: A learned multi-scale
representation for scene text recognition,” in Proceedings of the IEEE
Conference on Computer Vision and Pattern Recognition , 2014, pp.
4042–4049.
[18] K. Wang and S. Belongie, “Word spotting in the wild,” in European
Conference on Computer Vision . Springer, 2010, pp. 591–604.
[19] B. Shi, X. Bai, and C. Yao, “An end-to-end trainable neural network
for image-based sequence recognition and its application to scene
text recognition,” IEEE transactions on pattern analysis and machine
intelligence , vol. 39, no. 11, pp. 2298–2304, 2016.
[20] P. He, W. Huang, Y . Qiao, C. C. Loy, and X. Tang, “Reading scene
text in deep convolutional sequences,” in Thirtieth AAAI conference on
artificial intelligence , 2016.
[21] X. Yang, D. He, Z. Zhou, D. Kifer, and C. L. Giles, “Learning to read
irregular text with attention mechanisms.” in IJCAI , vol. 1, no. 2, 2017,
p. 3.
[22] W. Liu, C. Chen, and K.-Y . K. Wong, “Char-net: A character-aware
neural network for distorted scene text recognition,” in Thirty-Second
AAAI Conference on Artificial Intelligence , 2018.
[23] F. Zhan and S. Lu, “Esir: End-to-end scene text recognition via iter-
ative image rectification,” in Proceedings of the IEEE Conference on
Computer Vision and Pattern Recognition , 2019, pp. 2059–2068.
[24] J. L. Ba, J. R. Kiros, and G. E. Hinton, “Layer normalization,” arXiv
preprint arXiv:1607.06450 , 2016.
[25] K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image
recognition,” in Proceedings of the IEEE conference on computer vision
and pattern recognition , 2016, pp. 770–778.
[26] T.-Y . Lin, P. Goyal, R. Girshick, K. He, and P. Doll ´ar, “Focal loss
for dense object detection,” in Proceedings of the IEEE international
conference on computer vision , 2017, pp. 2980–2988.
[27] A. Mishra, K. Alahari, and C. Jawahar, “Top-down and bottom-up cues
for scene text recognition,” in 2012 IEEE Conference on Computer
Vision and Pattern Recognition . IEEE, 2012, pp. 2687–2694.[28] S. M. Lucas, A. Panaretos, L. Sosa, A. Tang, S. Wong, R. Young,
K. Ashida, H. Nagai, M. Okamoto, H. Yamamoto et al. , “Icdar 2003
robust reading competitions: entries, results, and future directions,”
International Journal of Document Analysis and Recognition (IJDAR) ,
vol. 7, no. 2-3, pp. 105–122, 2005.
[29] D. Karatzas, F. Shafait, S. Uchida, M. Iwamura, L. G. i Bigorda, S. R.
Mestre, J. Mas, D. F. Mota, J. A. Almazan, and L. P. De Las Heras,
“Icdar 2013 robust reading competition,” in 2013 12th International
Conference on Document Analysis and Recognition . IEEE, 2013, pp.
1484–1493.
[30] D. Karatzas, L. Gomez-Bigorda, A. Nicolaou, S. Ghosh, A. Bagdanov,
M. Iwamura, J. Matas, L. Neumann, V . R. Chandrasekhar, S. Lu et al. ,
“Icdar 2015 competition on robust reading,” in 2015 13th International
Conference on Document Analysis and Recognition (ICDAR) . IEEE,
2015, pp. 1156–1160.
[31] T. Quy Phan, P. Shivakumara, S. Tian, and C. Lim Tan, “Recognizing
text with perspective distortion in natural scenes,” in Proceedings of the
IEEE International Conference on Computer Vision , 2013, pp. 569–576.
[32] A. Risnumawan, P. Shivakumara, C. S. Chan, and C. L. Tan, “A robust
arbitrary text detection system for natural scene images,” Expert Systems
with Applications , vol. 41, no. 18, pp. 8027–8048, 2014.
[33] https://github.com/ICMLA-SAFL/SAFL pytorch.
[34] M. Jaderberg, K. Simonyan, A. Vedaldi, and A. Zisserman, “Deep
structured output learning for unconstrained text recognition,” arXiv
preprint arXiv:1412.5903 , 2014.
[35] Z. Cheng, F. Bai, Y . Xu, G. Zheng, S. Pu, and S. Zhou, “Focusing
attention: Towards accurate text recognition in natural images,” in
Proceedings of the IEEE international conference on computer vision ,
2017, pp. 5076–5084.
[36] Z. Cheng, Y . Xu, F. Bai, Y . Niu, S. Pu, and S. Zhou, “Aon: Towards
arbitrarily-oriented text recognition,” in Proceedings of the IEEE Con-
ference on Computer Vision and Pattern Recognition , 2018, pp. 5571–
5579.
[37] F. Bai, Z. Cheng, Y . Niu, S. Pu, and S. Zhou, “Edit probability for scene
text recognition,” in Proceedings of the IEEE Conference on Computer
Vision and Pattern Recognition , 2018, pp. 1508–1516.
[38] M. Liao, J. Zhang, Z. Wan, F. Xie, J. Liang, P. Lyu, C. Yao, and
X. Bai, “Scene text recognition from two-dimensional perspective,” in
Proceedings of the AAAI Conference on Artificial Intelligence , vol. 33,
2019, pp. 8714–8721.
[39] H. Li, P. Wang, C. Shen, and G. Zhang, “Show, attend and read: A simple
and strong baseline for irregular text recognition,” in Proceedings of the
AAAI Conference on Artificial Intelligence , vol. 33, 2019, pp. 8610–
8617.