MobText: A Compact Method for Scene Text Localization
Luis Gustavo Lorgus Decker
1, a
, Allan da Silva Pinto
, Jose Luis Flores Campana
Manuel Cordova Neira
, Andreza A. dos Santos
, Jhonatas S. Conceic¸
, Marcus A. Angeloni
Lin Tzy Li
and Ricardo da S. Torres
RECOD Lab., Institute of Computing, University of Campinas, 13083-852, Brazil
AI R&D Lab, Samsung R&D Institute Brazil, 13097-160, Brazil
Department of ICT and Natural Sciences, Norwegian University of Science and Technology (NTNU),
Alesund, Norway,,, {manuel.diduchi, andi.apsantos,
jhonatassantosdejesus17}, {m.angeloni,},
Scene Text Detection, Mobile Devices, Object Detector Networks, MobileNetV2, Single Shot Detector.
Multiple research initiatives have been reported to yield highly effective results for the text detection problem.
However, most of those solutions are very costly, which hamper their use in several applications that rely
on the use of devices with restrictive processing power, like smartwatches and mobile phones. In this paper,
we address this issue by investigating the use of efficient object detection networks for this problem. We
propose the combination of two light architectures, MobileNetV2 and Single Shot Detector (SSD), for the
text detection problem. Experimental results in the ICDAR’11 and ICDAR’13 datasets demonstrate that our
solution yields the best trade-off between effectiveness and efficiency and also achieved the state-of-the-art
results in the ICDAR’11 dataset with an f-measure of 96.09%.
Reading text in images is still an open problem in
computer vision and image understanding research
fields. In fact, this problem has attracted a lot of at-
tention of these communities due to large number of
modern applications that can potentially benefit from
this knowledge, such as self-driving vehicles (Yan
et al., 2018; Zhu et al., 2018), robot navigation, scene
understanding (Wang et al., 2018), assistive technolo-
gies (Yi et al., 2014), among others.
Several methods have been recently proposed in
the literature towards localizing textual information
in scene images. In general, the text reading prob-
lem is divided into two separated tasks, localization
and recognition, in which the former seeks to localize
delimited candidate regions that contain textual infor-
mation, while the second is responsible for recogniz-
ing the text inside the candidate regions found during
Part of results presented in this work were ob-
tained through the “Algoritmos para Detecc¸
ao e Reconhec-
imento de Texto Multil
ıngue” project, funded by Samsung
onica da Amaz
onia Ltda., under the Brazilian Infor-
matics Law 8.248/91.
Figure 1: Examples of textual elements with different font
sizes and styles.
localization task. In both tasks, the inherent variabil-
ity of a text (e.g., size, color, font style, background
clutter, and perspective distortions), as illustrated in
Fig. 1, makes text reading a very challenging prob-
Among the approaches for localizing texts in im-
ages, the deep-learning-based techniques are the most
promising strategy to reach high detection accuracy.
He et al., for example, presented a novel technique
Non-maximum suppression
Convolutional layers
Figure 2: Overview of the proposed method for text local-
for scene text detection by proposing a Convolutional
Neural Network (CNN) architecture (He et al., 2016)
that focuses on extracting text-related regions and
specific characteristics of text. The authors intro-
duced a deep multi-task learning mechanism to train
the Text-CNN efficiently, where each level of the su-
pervised information (text/non-text label, character
label, and character mask) is formulated as a learn-
ing task, besides a pre-processing method which en-
hances the contrast of small-size region improving the
local stability of text regions.
Although the proposed CNN presented a reason-
able efficiency in detecting candidate regions, with
a processing time of about 0.5 seconds per image,
the pre-processing step requires about 4.1 seconds per
image, which may prevent a real-time detection.
Another venue that may render outstanding re-
sults in terms of effectiveness consists in combining
different deep learning architectures to benefit from
complementary information to make a better deci-
sion. In this vein, (Zhang et al., 2016) introduced
an approach based on two Fully Convolutional Net-
work (FCN) architectures for predicting a salient map
of text regions in a holistic manner (named as Text-
Block FCN), and also for predicting the centroid of
each character. Similarly, Tang et al. (Tang and Wu,
2017) proposed an ensemble of three modified VGG-
16 networks: the first extracts candidate text regions
(CTR); the second network refines the coarse CTR de-
tected by the first model, segmenting them into text;
and finally, the refined CTR are served to a classifi-
cation network to filter non-text regions and obtain
the final text regions. The CTR extractor network is
a modified VGG-16 that, in the training process, re-
ceives the edges of the text as supervisory informa-
tion in the first blocks of convolutional layers and
the segmented text regions in the last blocks. Both
strategies present issues in terms of computational ef-
ficiency that could make their use unfeasible in re-
strictive computing scenarios (e.g., mobile devices).
Towards having a truthfully single stage text de-
tection, Liao et al. (Liao et al., 2018) proposed an
end-to-end solution named TextBoxes++, which han-
dles arbitrary orientation of word bounding boxes,
whose architecture inherits from the VGG-16. Sim-
ilarly to TextBoxes++, Zhu et al. proposed a deep
learning approach (Zhu et al., 2018) also based on
the VGG-16 architecture, but for detecting text-based
traffic sign. Both techniques presented outstanding
detection rates, though rely on the VGG-16 archi-
tecture, which could be considered inadequate for
restrictive computing scenarios due its model size.
In contrast, lighter CNN architectures, such as Mo-
bileNet (Howard et al., 2017), present a very compet-
itive alternative for this scenario, with a model size
of 4.2 millions of parameters and the FLOPS of 569
millions, for instance.
With those remarks, we propose a novel method
for text localization considering efficiency and effec-
tiveness trade-offs. Our approach combines two light
architectures that were originally proposed for object
detection MobileNetV2 (Sandler et al., 2018) and
SSD (Liu et al., 2016) – and adapts them to our prob-
lem. The main contributions of this paper are: (i) the
proposal of an effective method for text localization
task in scene images, which presented better or com-
petitive results (when compared with state-of-the-art
methods) at a low computational costs in terms of
model size and processing time; and (ii) a compara-
tive study, in the context of text localization, compris-
ing widely used CNN architectures recently proposed
for object detection.
Fig. 2 illustrates the overall framework of our ap-
proach for text localization, which uses MobileNetV2
as feature extractor and then SSD (convolutional lay-
ers) as multiple text bonding boxes detector. Next,
We will detail the CNN architectures used, then ex-
plain the learning mechanism adopted for finding a
proper CNN model for the problem.
2.1 Characterization of Text Regions
with MobileNetV2
The MobileNetV2 is a new CNN specifically de-
signed for restrictive computing environments that in-
cludes two main mechanisms for decreasing the mem-
ory footprints and the number of operations while
keeping the effectiveness of its precursor architecture,
the MobileNet (Sandler et al., 2018): the linear bot-
tlenecks and the inverted residuals.
Fig. 3 shows the MobileNetV2 architecture used
to characterize text candidate regions. The bottleneck
x2 x3
conv. layer (3x3)
channels (32)
stride (2)
conv. layer (1x1)
channels (1280)
stride (1)
residual block
channels (16)
stride (1)
residual block
channels (24)
stride (2)
residual block
channels (32)
stride (2)
residual block
channels (64)
stride (2)
residual block
channels (96)
stride (2)
residual block
channels (160)
stride (2)
residual block
channels (320)
stride (2)
Figure 3: MobileNetV2 architecture used in this work and
its parameters. More detail on Bottlenet residual block
in (Sandler et al., 2018).
residual block implements the optimization mecha-
nisms aforementioned considering the convolutional
operations with a kernel of size 3 × 3. The first bot-
tleneck block uses an expansion factor of 1, while the
remaining blocks use an expansion factor of 6, as sug-
gested by Sandler et al. (Sandler et al., 2018).
2.2 Detecting Multiple Text Bounding
Boxes via SSD
The localization of text regions in scene images is
challenging due to inherent variability of the text,
such as size, color, font style, and distortions. The
text localization should handle multiple scales and
bounding boxes with varying aspect ratio. Although
several authors consider the image pyramid for per-
forming multi-scale detection, it is quite costly, which
may be impractical in a restrictive computing sce-
nario. Thus, we use the Single Shot detector (SSD)
framework (Liu et al., 2016), a state-of-the-art method
for object detection. The SSD approach includes a
feature pyramid mechanism that allows the identifica-
tion of text regions in multiple scales. Specifically, in
the framework, the authors adopt a top-down fusion
strategy to build new features with strong semantics
while keeping fine details. Text detections are per-
formed based on multiple new constructed features
respectively during a single forward pass. All detec-
tion results from each layer are refined by means of a
non-maximum suppression (NMS) process (Neubeck
and Gool, 2006).
2.3 Using Linear Bottlenecks and
Inverted Residuals Bottlenecks for
Memory Efficiency
Besides the use of depthwise separable convolutions
operations, MobileNetV2 introduced linear bottle-
necks in the convolutional blocks. This reduces
the number of parameters of a neural network and
captures the low-dimensional subspace, supposing
that such low-dimensional subspace is embedded in
a manifold formed by a set of activation tensors.
In (Sandler et al., 2018), Sandler et al. showed empir-
ical evidences that the use of linear layers is impor-
tant to prevent non-linearity added from destroying
information. Experiments conducted by the authors
showed that non-linear bottlenecks, built with recti-
fied linear units, can decrease the performance sig-
nificantly in comparison with linear bottlenecks. By
using the idea of Inverted Residual bottlenecks, the
authors achieved better memory use, reducing a sig-
nificant amount of computation. We follow this idea
in this paper.
2.4 Learning
The main decisions we took in the learning phase of
our network are described below.
Objective Function. Similar to (Liu et al., 2016),
we use a multi-task loss function to learn the bound-
ing boxes locations and text/non-text predictions
(Eq. 1). Specifically, x
i j
indicates a match (x
i j
= 1)
or non-match (x
i j
= 0) between i-th default bound-
ing boxes, j-th ground-truth bounding boxes; N is the
number of matches; and the α parameter is used to
weight the localization loss (L
) and the confidence
loss (L
con f
). The used loss function can be defined as:
L(x, c, l, g) =
con f
(x, c) + αL
(x, l, g)) (1)
We adopted the smooth L1 loss (Girshick, 2015)
for L
between the predicted box (l) and the ground
truth box (g), and a sigmoid function for L
con f
. Plus,
we consider α = 1 in the same fashion as (Girshick,
Hard Example Mining. The hard example miner
is a mechanism used to prevent imbalances between
negative and positive examples in the training phase.
During the search for text during the training, we usu-
ally have several non-text bounding boxes and few
text bounding boxes. To mitigate the training with im-
balanced data, we sort the negative bounding boxes
according to their confidence, selecting the negative
samples with higher confidence value, considering a
ratio proportion of 3:1 with the positive samples.
This section presents the datasets, metrics, and proto-
cols used for evaluating the proposed method.
3.1 Datasets
We evaluated the proposed methods in two datasets
widely used for evaluating text localization methods:
ICDAR’11Karatzas et al. (2011), that contains 551
digitally created images, such as headers, logos, cap-
tions, among others, and ICDAR’13Karatzas et al.
(2013), containing 462 born-digital or scene text im-
ages (captured under a wide variety, such as blur,
varying distance to camera, font style and sizes, color,
texture, etc). We also used the SynthTextGupta et al.
(2016) dataset to help training our network due to
the small size of the ICDAR’s datasets. We have not
used ICDAR’15Karatzas et al. (2015) and some other
newer datasets, because our method is not tailored to
the prediction of oriented bounding boxes, which is
required to handle such multioriented datasets.
3.2 Evaluation Metrics
Effectiveness. We evaluated the effectiveness of the
methods in terms of recall, precision, and f-measure.
Here, we consider a correct detection (true positive)
if the overlap between the ground-truth annotation
and detected bounding box, which is measured by
computing the intersection over union, is greater than
50%. Otherwise, the detected bounding box is con-
sidered an incorrect detection.
Efficiency. The efficiency aspects considered both
the processing time and the disk usage (in MB). We
used the GNU/Linux time command to measure the
processing time, while the disk usage considered the
size of the learned models. All experiments were
performed considering a Intel(R) Core(TM) i7-8700
CPU @ 3.20GHz with 12 cores, a Nvidia GTX 1080
TI GPU, and 64GB of RAM.
3.3 Evaluation Protocols
The experiments were divided into three steps: train-
ing, fine-tuning, and test. For the training step, we
used four subsets of the SynthText dataset. This
dataset comprises of images with synthetic texts
added in different backgrounds and we selected
samples of the dataset considering 10 (9.25%), 20
(18.48%), and 30 (27.71%) images per background,
then finally the whole dataset. The resulting sub-
sets were again divided into train and validation, us-
ing 70% for training and 30% for validation. Using
these collections, we trained a model with random
initialization parameters for 30 epochs. For the fine-
tuning step, we took the model trained in SynthText
and continued this training using ICDAR’11 or IC-
DAR’13 training subsets, stopping when we reached
2000 epochs. The number of epochs was defined em-
pirically. Finally, for the test step, we evaluated each
fine-tuned model in the test subset of ICDAR’11 or
Experimental Setup. We conducted the training of
the proposed method considering a single-scale in-
put, and therefore, all input images were resized to
300 × 300 pixels. The training phase was performed
using a batch size of 24 and we used the RMSprop op-
timizer (Tieleman and Hinton, 2012) with a learning
rate of 4 × 10
. We also use the regularization L2-
norm, with a λ = 4 × 10
, to prevent possible over-
3.4 State-of-the-Art Object Detection
Methods for Comparison
This section provides an overview of the chosen meth-
ods for comparison purpose. For a fair compari-
son, we selected recent approaches specifically de-
signed for a fast detection, including SqueezeDet and
YOLOv3. We also use methods for text localiza-
tion that presents good compromise among effective-
ness and efficiency as baselines, which are briefly de-
scribed in this section.
TextBoxes. This method consists of a Fully Con-
volutional Network (FCN) adapted for text detection
and recognition (Liao et al., 2017). This network uses
the VGG-16 network as feature extractor followed by
multiple output layers (text-boxes layers), similar to
SSD network. At the end, the Non-maximum sup-
pression (NMS) process is applied to the aggregated
outputs of all text-box layers.
TextBoxes++. This method extends the TextBoxes
method (Liao et al., 2017) toward detecting arbitrary-
oriented text, instead of only (near)-horizontal bound-
ing boxes (Liao et al., 2018). TextBoxes++ also
brings improvements in the training phase, which
leads to a further performance boost, in terms of accu-
racy, especially for detecting texts in multiple scales.
In this work, the authors combine the detection scores
of CRNN recognition method (Shi et al., 2017) with
the TextBoxes++ to improve the localization results
and also to have an end-to-end solution.
SSTD. Single-shot text detector proposed
by He et al. (He et al., 2017) designed a natural
scene text detector that directly outputs word-level
bounding boxes without post-processing, except for
a simple NMS. The detector can be decomposed
into three parts: a convolutional component, a text-
specific component, and a box prediction component.
The convolutional and box prediction components
are inherited from the SSD detector (Liu et al., 2016)
and the authors proposed a text-specific component
which consists of a text attention module and a
hierarchical inception module.
SqueezeDet. This network was proposed to detect
objects for the autonomous driving problem, which
requires a real-time detection (Wu et al., 2017).
The SqueezeDet contains a single-stage detection
pipeline, which comprises three components: (i) a
FCN responsible for generating the feature map for
the input images; (ii) a convolutional layer responsi-
ble for detecting, localizing, and classifying objects
at the same time; and (iii) the non-maximum suppres-
sion (NMS) method, which is applied to remove the
overlapped bounding boxes.
YOLOv3. This is a convolutional network origi-
nally proposed for the object detection problem (Red-
mon and Farhadi, 2018). Similarly to SSD network,
the YOLOv3 predicts bounding boxes and class prob-
abilities, at the same time.
This section presents the experimental results of the
proposed method (SSD-MobilenetV2) and a compar-
ison with the state-of-the-art methods for text local-
ization. Table 1 shows the results for the evaluated
methods considering the ICDAR’11 dataset. In this
case, the SSD-MobilenetV2 method achieved the best
results with Precision, Recall, and F-measure values
of 97.40%, 94.81%, and 96.09%, respectively. On
the other hand, the SqueezeDet network presented the
lowest Precision and F-measure among the evaluated
methods (56.36% and 66.01%, respectively). In turn,
the TextBoxes achieved the lowest results of Recall
With regard to ICDAR’13 dataset, the SSTD
methods presented the highest Recall (82.19%), and
F-measure (86.33%), while the YOLOv3 reached the
best results in terms of Precision (Table 1). Note,
however, that the SSD-MobileNetV2 yields very
competitive results for this dataset as well, in terms
of Precision.
As we could observe, the proposed approach pre-
sented some difficult in localizing scene text in the IC-
DAR’13 dataset. In comparison with results achieved
Figure 4: Comparison results among the evaluated methods
considering aspects of efficacy and efficiency.
Figure 5: Two high resolution examples of ICDAR’13
dataset with both medium-sized text (detected by our
method) and small-sized (not detected).
for the ICDAR’11, the precision and recall rates de-
creased 9.36 and 31.61 percentage points, respec-
tively, which suggest that our network did not local-
ized several candidate regions containing texts.
To understand the reasons that led the proposed
method to have this difficult in localizing text for
the ICDAR’13 datasets, we performed an analysis of
failure cases taking into account the relative area of
missed bounding boxes. Fig. 4 presents a box-plot
graph that shows the distribution of the relative area
of bounding boxes (i.e., ratio of bounding box area to
Table 1: Comparison of effectiveness among the evaluated deep learning-based methods for the ICDAR’11 and ICDAR13
Datasets ICDAR’11 ICDAR’13
Methods P (%) R (%) F (%) P (%) R (%) F (%)
SSD-MobilenetV2 97.40 94.81 96.09 88.38 66.67 76.00
SSTD 89.28 78.53 83.56 90.91 82.19 86.33
TextBoxes 92.15 71.93 80.80 88.84 74.16 80.83
TextBoxes++ 95.76 90.51 93.06 90.49 80.82 85.38
YOLOv3 94.27 89.21 91.67 92.01 75.71 83.07
SqueezeDet 56.36 79.66 66.01 29.41 62.47 39.99
Figure 6: Comparison among distributions of relative areas
of bounding boxes from Ground-Truth (GT), False Nega-
tives (FN) cases, and False Positive (FP) cases. We omitted
the points considered outliers for a better visualization.
image area) for the ground-truth, false positive cases,
and false negative cases.
As we can observe, the missed bounding boxes
(false negative cases) have a small relative area. More
precisely, 75% of false negative cases (third quartile
of FN box-plot) have a relative area up to 0.01 and
correspond to 50% of the bounding box present in the
ground-truth (median of GT box-plot). This results
suggest to us that high resolution images with rela-
tively small text (see Fig. 5) are specially challenging
to our method. To overcome this limitation, future
investigations can be conducted to devise an architec-
ture to better localize bounding boxes with multiple
scales such as Feature Pyramid Networks (FPNs), as
proposed by (Lin et al., 2017).
In term of the efficiency of the presented methods,
Fig. 4 summarizes the results considering the metrics
used to assess the effectiveness of the evaluated meth-
ods, in terms of F-measure, along with the metrics for
measuring the efficiency of those methods, consider-
ing the ICDAR’11 and ICDAR’13 datasets.
Regarding the efficiency (processing time and disk
usage), the proposed method (SSD-MobilenetV2)
yielded very competitive results, taking only 0.45
and 0.55 seconds per image, considering the IC-
DAR’11 and ICDAR’13, respectively. Compar-
ing SSD-MobilenetV2 with the baseline methods
originally proposed for text localization (TextBoxes,
TextBoxes++, SSTD), the proposed method presented
the very competitive results with a processing time
of 0.67 seconds per image and with disk usage of
about 37.0MB. In contrast, the most effective baseline
methods, the SSTD and TextBoxes++ networks, pre-
sented competitive and worse results in terms of ef-
fectiveness and processing time, respectively, in com-
parison with the proposed method. Regarding the disk
usage, the SSD-MobileNetV2 also presented the best
balance between accuracy and model size.
Now, when compared with the state-of-the-art ap-
proaches for object detection, the proposed method
also presented competitive results. In this case,
the fastest approach for text localization was the
SqueezeDet network, which takes about 0.1 seconds
per image, on average. However, when we take
into account the trade-off between efficiency and ef-
fectiveness, we can safely argue that the proposed
method presented a better compromise between these
two measures. Fig. 7 and Fig. 8 provide some cases of
success (first column) and of failure of the proposed
method for the ICDAR’11 and ICDAR’13 datasets.
For the first (see Fig), the proposed method was able
to localize textual elements with different font styles
and even multi-oriented texts. For the latter, the pro-
posed method was able to localize text in several con-
texts such as in airport signs, traffic signs, text in ob-
jects, among others. Failures are due to compression
artifacts, the high similarity between the background
and the text colors, small texts, and lighting condi-
How to perform efficient and effective text detec-
tion in scene images in restrictive computing envi-
ronments? To address that research problem, we pre-
sented a new method based on the combination of two
Figure 7: Examples of success (first row) and failure (second row) cases of the proposed approach for the ICDAR’11 dataset.
Green bounding boxes indicate the regions correctly localized (true positives cases), while red bounding boxes show candidate
regions were not detected by our method (false negatives cases).
Figure 8: Examples of success (first row) and failure (second row) cases of the proposed approach for the ICDAR’13 dataset.
Green bounding boxes indicate the regions correctly localized (true positives cases), while red bounding boxes show candidate
regions were not detected by our method (false negatives cases).
light architectures, MobileNetV2 and Single Shot De-
tector (SSD), which yielded better or comparable ef-
fectiveness performance when compared with state-
of-the-art baselines despite having a low processing
time and small model size. Compared with other
object detector solutions, our methods is the most
promising. Our findings disagree with the discussion
provided in (Ye and Doermann, 2015), as we demon-
strated that adapting object detector networks for text
detection is a promising research venue.
Future research efforts will focus on better char-
acterizing both small and large candidate regions to
localize text in multiple scales such as Feature Pyra-
mid Networks.
Girshick, R. (2015). Fast r-cnn. In The IEEE International
Conference on Computer Vision (ICCV).
Gupta, A., Vedaldi, A., and Zisserman, A. (2016). Synthetic
data for text localisation in natural images. In 2016 IEEE
Conference on Computer Vision and Pattern Recognition
(CVPR), pages 2315–2324.
He, P., Huang, W., He, T., Zhu, Q., Qiao, Y., and Li, X.
(2017). Single shot text detector with regional attention.
In 2017 IEEE International Conference on Computer Vi-
sion (ICCV), pages 3066–3074.
He, T., Huang, W., Qiao, Y., and Yao, J. (2016). Text-
attentional convolutional neural network for scene text
detection. IEEE Transactions on Image Processing,
Howard, A. G., Zhu, M., Chen, B., Kalenichenko, D.,
Wang, W., Weyand, T., Andreetto, M., and Adam,
H. (2017). Mobilenets: Efficient convolutional neu-
ral networks for mobile vision applications. CoRR,
Karatzas, D., Gomez-Bigorda, L., Nicolaou, A., Ghosh, S.,
Bagdanov, A., Iwamura, M., Matas, J., Neumann, L.,
Chandrasekhar, V. R., Lu, S., Shafait, F., Uchida, S., and
Valveny, E. (2015). Icdar 2015 competition on robust
reading. In 2015 13th International Conference on Doc-
ument Analysis and Recognition (ICDAR), pages 1156–
Karatzas, D., Mestre, S. R., Mas, J., Nourbakhsh, F., and
Roy, P. P. (2011). Icdar 2011 robust reading competition
- challenge 1: Reading text in born-digital images (web
and email). In 2011 International Conference on Docu-
ment Analysis and Recognition, pages 1485–1490.
Karatzas, D., Shafait, F., Uchida, S., Iwamura, M., Big-
orda, L. G. i., Mestre, S. R., Mas, J., Mota, D. F., Al-
an, J. A., and de las Heras, L. P. (2013). Icdar
2013 robust reading competition. In Proceedings of the
2013 12th International Conference on Document Anal-
ysis and Recognition, ICDAR ’13, pages 1484–1493,
Washington, DC, USA.
Liao, M., Shi, B., and Bai, X. (2018). Textboxes++: A
single-shot oriented scene text detector. IEEE Transac-
tions on Image Processing, 27(8):3676–3690.
Liao, M., Shi, B., Bai, X., Wang, X., and Liu, W. (2017).
Textboxes: A fast text detector with a single deep neural
network. In Proceedings of the Thirty-First AAAI Confer-
ence on Artificial Intelligence, February 4-9, 2017, San
Francisco, California, USA., pages 4161–4167.
Lin, T., Doll
ar, P., Girshick, R., He, K., Hariharan, B., and
Belongie, S. (2017). Feature pyramid networks for object
detection. In 2017 IEEE Conference on Computer Vision
and Pattern Recognition (CVPR), pages 936–944.
Liu, W., Anguelov, D., Erhan, D., Szegedy, C., Reed, S., Fu,
C.-Y., and Berg, A. C. (2016). SSD: Single shot multibox
detector. In Leibe, B., Matas, J., Sebe, N., and Welling,
M., editors, Computer Vision ECCV 2016, pages 21–
37, Cham. Springer International Publishing.
Neubeck, A. and Gool, L. V. (2006). Efficient non-
maximum suppression. In 18th International Conference
on Pattern Recognition (ICPR’06), volume 3, pages 850–
Redmon, J. and Farhadi, A. (2018). Yolov3: An incremental
improvement. CoRR, abs/1804.02767.
Sandler, M., Howard, A., Zhu, M., Zhmoginov, A., and
Chen, L. (2018). Mobilenetv2: Inverted residuals and
linear bottlenecks. In 2018 IEEE/CVF Conference on
Computer Vision and Pattern Recognition, pages 4510–
Shi, B., Bai, X., and Yao, C. (2017). An end-to-end train-
able neural network for image-based sequence recogni-
tion and its application to scene text recognition. IEEE
Transactions on Pattern Analysis and Machine Intelli-
gence, 39(11):2298–2304.
Tang, Y. and Wu, X. (2017). Scene text detection
and segmentation based on cascaded convolution neu-
ral networks. IEEE Transactions on Image Processing,
Tieleman, T. and Hinton, G. (2012). Lecture 6.5-rmsprop:
Divide the gradient by a running average of its recent
magnitude. COURSERA: Neural networks for machine
learning, 4(2):26–31.
Wang, L., Wang, Z., Qiao, Y., and Van Gool, L. (2018).
Transferring deep object and scene representations for
event recognition in still images. International Journal
of Computer Vision, 126(2):390–409.
Wu, B., Iandola, F., Jin, P. H., and Keutzer, K. (2017).
Squeezedet: Unified, small, low power fully convo-
lutional neural networks for real-time object detection
for autonomous driving. In 2017 IEEE Conference on
Computer Vision and Pattern Recognition Workshops
(CVPRW), pages 446–454.
Yan, C., Xie, H., Liu, S., Yin, J., Zhang, Y., and Dai,
Q. (2018). Effective uyghur language text detection in
complex background images for traffic prompt identifi-
cation. IEEE Trans. Intelligent Transportation Systems,
Ye, Q. and Doermann, D. S. (2015). Text detection and
recognition in imagery: A survey. IEEE Trans. Pattern
Anal. Mach. Intell., 37(7):1480–1500.
Yi, C., Tian, Y., and Arditi, A. (2014). Portable camera-
based assistive text and product label reading from hand-
held objects for blind persons. IEEE/ASME Transactions
on Mechatronics, 19(3):808–817.
Zhang, Z., Zhang, C., Shen, W., Yao, C., Liu, W., and Bai,
X. (2016). Multi-oriented text detection with fully con-
volutional networks. In The IEEE Conference on Com-
puter Vision and Pattern Recognition (CVPR).
Zhu, Y., Liao, M., Yang, M., and Liu, W. (2018). Cascaded
segmentation-detection networks for text-based traffic
sign detection. IEEE Transactions on Intelligent Trans-
portation Systems, 19(1):209–219.