Text and Object Recognition using Deep Learning for Visually Impaired People

International Journal of Trend in Scientific Research and Development (IJTSRD)
Volume 4 Issue 5, July-August 2020 Available Online: www.ijtsrd.com e-ISSN: 2456 – 6470
@ IJTSRD | Unique Paper ID – IJTSRD31508 | Volume – 4 | Issue – 5 | July-August 2020 Page 624
Text and Object Recognition using
Deep Learning for Visually Impaired People
R. Soniya1, B. Mounica1, A. Joshpin Shyamala1, Mr. D. Balakumaran2
2Assistant Professor, M.E,
1,2Department of Electronics and Communication Engineering,
1,2S. A. Engineering College, Chennai, Tamil Nadu, India
ABSTRACT
The main aim of this paper is to aid the visually impaired people with object
detection and text detection using deep learning. Object detection is done
using a convolution neural network and text recognition is done by optical
character recognition. The detected output is converted into speechusingtext
to the speech synthesizer. Object detection comprises of two methods. One is
object localization and the other is image classification. Image classification
refers to the prediction of classes of different objects within an image. Object
localization infers the location of objects using bounding boxes.
KEYWORDS: object detection, CNN, text detection, OCR, TTS
How to cite this paper: R. Soniya | B.
Mounica | A. Joshpin Shyamala | Mr. D.
Balakumaran "Text and Object
Recognition using Deep Learning for
Visually Impaired
People" Published in
International Journal
of Trend in Scientific
Research and
Development(ijtsrd),
ISSN: 2456-6470,
Volume-4 | Issue-5,
August 2020, pp.624-628, URL:
www.ijtsrd.com/papers/ijtsrd31508.pdf
Copyright © 2020 by author(s) and
International Journal of TrendinScientific
Research and Development Journal. This
is an Open Access
article distributed
under the terms of
the Creative Commons Attribution
License (CC BY 4.0)
(http://creativecommons.org/licenses/by
/4.0)
I. INTRODUCTION
Although much advancement has been made in the image
processing field, extracting objects from the color image is
not an easy task. This paper dealswith taking out the objects
and text alone from a frame. The object is recognized by
passing theinput video toa trainedmodelwhich leads to the
classification of objects between variousclasses.Themodelis
trained by MS-COCO (Common Objects in Context) [1] and
further tuned finely by using PASCAL VOC0712 [2] has some
pre-trained weights. It makes use of single-shot detectors
(SSD)and mobile nets result in fast and real-time object
detection.
Figure 1:Overall diagram for detection and recognition
A Text detectionmodelisbased ona fullyconvolution neural
network [2][3]. Itwas adopted tocentralize the text regions.
Generally, the feature of the actual input image is first
extracted. The one convolution is applied to output dense
per-pixel predictions of text to be present. For each positive
sample, the channel assumes a probability. As referred in
figure:1 The preprocessed image input is forwarded to the
neural network.
There it compares the input with the trained dataset after
detecting the object and text. It generates output as to what
object or text is that. This is followed by text to speech
conversion process to help the visually impaired person.
II. CONVOLUTION NEURAL NETWORK
CNN [3] is primarily used for pattern recognition within the
image, thus making it suitable for the image-focused taskby
reducing parameters to set up a model. It consists of three
layers namely, convolution layer, pooling, fully connected
layer. Figure 1: Shows simplified CNN architecture
comprising of feature extraction and classification.
Convolution layer performs convolution operation between
an array of input pixels and kernel of a particular size which
slides all over the input frame
IJTSRD31508

International Journal of Trend in Scientific Research and Development (IJTSRD) @ www.ijtsrd.com eISSN: 2456-6470
Figure 2: Simplified CNN Architecture
The resulting output is passed through the activation
function (Rectified linear unit) in-order to removenonlinear
functionalities figure (2). The Pooling layer simplifies the
overall dimension of the image by performing the down
sampling operation. It operates over each input activation
map and scales using the “MAX” function. Most of the CNN
contains a max-pooling layer with a kernel of 2x2 size.
Figure 3: Rectified linear unit (ReLU) activation
function
The output from the final pooling or convolution layer is
flattened and fedas input toa fully connected layer. The final
layer uses soft max activation function which is used to
classify based on probabilities obtained.
III. CAFFE OBJECT DETECTION MODELS
Caffe [4] (convolution Architecture for fast feature
embedding) is a deep learning framework that supports
different architecture for image classification and
segmentation. The following are the model used for object
recognition.
A. Single-shot detectors:
It is the combination of Yolo’s regression and a faster RCNN
mechanism [5]. It simplifies the complexity of the neural
network. The local feature extraction method is effective in
SSD where multiscale feature extraction. The recognition
speed of SSD is 59 fps (frames per second). It quenches the
need for real-timeimplementations. The only demeritisthat
its fragile detection of tiny objects. At the time of prediction,
the network generates a default box and adjusts it to match
with the object shape [6],[7].
Figure 3:localization and confidence using multiple
bounding boxes
B. Mobile Nets
While developing object detection networks we generally
utilize existing network architecture, suchas VGG or ResNet
The limitationis thatthese networkarchitecturescanbevery
huge in the order of 200-500MB. So it doesn't support
resource-constrained systems. Mobile nets differ from
traditional CNN through the usage of depth-wise separable
convolution (figure 4)[7],[8].
Figure 4: (Left) Standard convolutional layer with
batch normalization and ReLU. (Right) Depth- wise
separable convolution with depth-wise and pointwise
layers followed by batch normalization and ReLU
(figure and caption from Liu et al.).
Here, convolution is divided into two stages of 3x3 depth
wise convolutions followed bya 1x1 point wise convolution.
IV. TEXT RECOGNITION
Text recognition utilizes the help of OCR to detect the text
from an image. OCR is a popular model use to get the
character from the text. OCR is a self learning translator of
character in typed, printed or any documented and pictured
forms. It has various methods to recognize characters like
thinning thickening, feature extraction [9]. The major steps
corresponding to text detection comprises data collection,
pre processing, segmentation, features extracting and
classification. In case of first step, pre trained data sets or
own data sets can be used for the research. Pre processing
includes slant detection, slant correction, gray scale
conversion, normalization; noise removaletc.segmentationis
done to splits the words from sentences for better detection.
The feature extraction step extracts the features from the
separated words and compares it with features of trained
images to confirm with the matching classes. Finally
classifiers are used with respect to convenience for the
classification process.
Figure 5:Basic steps of OCR[10]

A. Data collection:
It involves acquiring the text from a frame in the form of
video of resolution using webcam [9]
B. Preprocessing:
Once the data is collected, various preprocessing techniques
are used to enhance and to improve the quality of the image.
It is an important stage prior to feature extraction because it
determines the quality of the image in the successive stages
[10].
C. Segmentation:
Segmentationisnothingbutseparating thevarious lettersof
a word. Only if there is a good segmentation, better
recognition rate can be obtained [11].
D. Feature extraction:
It is a stage where the required information is obtained by
using minimum number of resources. It also reduces the
overall dimensions of the image.
E. Classification:
It is process of categorizing a character into its suitable
category. There are different function to classify the
image[12],[13].
V. TEXT TO SPEECH SYNTHESIZER
Text to speech system (TTS) converts text into voice using a
synthesizer[14]which producesartificial speechofhuman.It
converts an arbitrary plain text into a corresponding
waveform.
The maincomponents of this systemare text processingand
speech generation [15]. The text processing unit is
responsible for producing a sequence of phonemic units
according to the input text and also includes text
normalization and tokenization. Then makes the text
into prosodic units, which is termed as phonetic (also called
as text-to-phoneme or grapheme-to-phoneme conversion).
These are then realized by speech generation system. A
general text to speech system is shown in figure 5[16]
Figure 5: A overview of TTS system
The Textnormalizationinvolvestransforming thewholetext
into standard format [17].Tokenization are used to split the
whole text intowords, keywords, phrasesand symbolsitalso
discards some punctuation like commas.
Speech synthesizer has to generate sounds to make those
words in input text. There are 40 phonemes available for 26
letters in English alphabet it is because some letters can be
read in multiple ways. But this may get harder for practical
implementation commonly called as prosody, in linguistics.
Within a word, a given phoneme may sounds differently
because the phonemes that come before and after it.
Thusgrapheme isthealternative approachwherewordsare
segmented into individual letters. Once the sequences of
words are converted into list of phonemes, then it will be
converted into speech. Following are the different speech
synthesis techniques.
A. Concatenative synthesis
It is a synthesizer that make use of preloaded record of
human voiceswhichisreferred toas units. Theunitscontains
lots of examples of different things, break the spoken
sentencesintowordsandthewordsinto phonemes[18].Thus
itcan easily rearrange tocreate newwordsandsentence.The
duration of units is defined and are in the range of 10 ms up
to 10 seconds.
B. formant synthesis
It is created by using additive synthesis and an acoustic
model instead of human speech input during runtime[19].
Factors like noise, frequencyandvoicearechangedwithtime
to produce signalsofartificial speech.Thisis often termedas
rules based synthesis. Most of the systems that use formant
synthesis have robotic or artificially generated voice which
are unique from human speech. It is trusted for its
intelligence and high speed, eliminating acoustic flaws that
usually infect concatenative systems. Visually impaired use
high speedsynthesizedspeech tonavigate throughcomputer
screensUsingscreen reader. Formant synthesizers normally
do not have a database of speech samples so, they are found
to be smaller programs than concatenative systems.
C. Articulatory synthesis
It is computational technique used for synthesizing speech
with respecttomodel ofhuman vocal tract(voiceprocess).It
is a mechanism to mimic a human voice[19].
VI. EXPERIMENTAL SETUP
A. Object Detection
To execute object detection using CNN it makes use of the
Caffe framework for deploying the object detection model.
The datasets include 20 classes [7], [20].
The model contains apre-trained. Caffe model file, many.
Proto txt files. Which has its parameterized structure within
the proto text file and also weights are predefined.
Thus a single frame of the input video is resized to the
dimension of 300x300(height & width) and forward passed
into the model. So that it can detect 19 objects in an image.
Multiple objects can be detected from a single frame.
B. trained dataset
The datasetcontainsthefollowingclassesincluding airplanes,
boats, bicycles, bottles, birds, buses, cars, cats, chairs, cows,
dining tables, dogs, horses, motorbikes, potted plants, people,
sofas, trains, and television monitors.
The probability during each detection is also checked. If the
confidence level is above the threshold level defined the
particularclass namewhich is referred toas label alongwith
their accuracy will be displayed and also a bounding box is
computed over the objects detected
And also the label detected will be created as a separate text
file which helps in producing the audio output of the object
detected.

A. labels and their accuracy during detection
TABLE I Different objects and text fonts that can be
detected
Since the labeling had to be done manually and also because
of the size of RAM we couldn’t produce enough data, or data
of assured quality, to properly train the neural network
model. We have reached the possible accuracy for a CNN
with images as input data, we have only created a model for
detecting 20 objects.
With the help of OCR it detects the different font styles with
approximate predictions of the word or sentence. Table
6.1(a) shows the list ofobjectsand Table 6.1(b) shows listof
text fonts which can be detected by the proposed algorithm.
Along with the detection italso produces voice output of the
detected object and text through the earphone connected to
the audio jack.
As far as the results are concerned the following are
observed to be achieved from the project. In case of text, itis
possible for the device to detect from either image inputs or
live reading of the texts. Itiscapable of recognizing standard
text formats of different fonts. Especially bold letters can be
detected fast and accurately. It is noted that words in light
background are recognized well with precision.
These results are based on the frame intensity and pixel
variations. The outcome is subject to vary in accuracy
depending on the camera specifications and focus done to
the object. The accuracy of the predicted objectsdifferdueto
different condition suchascontrast sensitivity, illumination,
noise and also it producesmaximumaccuracyif the objectis
closer to the camera in the range of 70 cm with good
lightning condition.
VII. CONCLUSION
This device iscapable ofacting as a smart device. It provides
visually impaired people with the list of objects in the
surrounding along with accuracy of prediction.Thereforethe
person will get to know about the hindrances in his path.
This is more like a voice assistant that tells name of the
things around us. Multiple objects in a single frame is
detected which is an added advantage.
REFERENCES
[1] T.-Y. Lin, J. Hays, M. Maire, P. Perona, S. Belongie, D.
Ramanan, L. Bourdev, L. Zitnick, R. Girshick and P.
Dollar, "MicrosoftCOCO: Common Objects in Context,"
arXiv:1405.0312v3, pp. 1-15, February 2015.
[2] X. Zhou et al., ”EAST: An efficient and accurate scene
text detector”, Proc-30th IEEEconf.Comput.vis. Pattern
Recoginition, CVPR 2017, vol.2017-janua, pp.2642-
2651, 2017.
[3] X. Liu, D. Liang, S. Yan, D. Chen, Y. Qiao and J. Yan,
”FOTS: Fast Oriented Text Spotting with a Unified
Network”, Proc. IEEEComput. Soc. Conf. Comput.Vis.
Pattern Recognit., pp.5676-5685, 2018.
[4] M. Everingham, L. V. Gool, C. K. I. Williams, J. Winn and
A. Zisserman, "The PASCAL Visual ObjectClasses(VOC)
Challenge," International Journal of Computer Vision,
vol. 88, no. 2, pp. 303-338, 2010. ****
[5] Krizhevsky, A., Sutskever, I., Hinton, G.E.: Imagenet
classiﬁcation with deep convo-lutional neural
networks. In: Advances in neural information
processing systems. pp. 1097–1105 (2012)****
[6] W. Liu, D. Anguelov, D. Erhan, C. Szegedy, S. Reed, C.-Y.
Fu, andA. C. Berg, “Ssd: Single shot multibox detector,”
in ECCV, 2016.(ssdku)***
[7] Liu, Wei, et al. "Ssd: Single shot multibox detector."
European conference on computer vision. Springer,
Cham, 2016.
[8] A. G. Howard, M. Zhu, B. Chen and D. Kalenichenko,
"MobileNets: Efficient ConvolutionalNeural Networks
for Mobile Vision," 2017.
[9] “mobilenet_ssd pretrained model”
https://github.com/chuanqi305/MobileNet-
SSD/blob/master/README.md
[10] A Survey on Optical Character Recognition System
Noman Islam, Zeeshan Islam, Nazia Noor, Journal of
Information & Communication Technology-JICTVol.10
Issue. 2, December 2016.
[11] Lund, W.B., Kennard, D.J., & Ringger, E.K. (2013).
Combining Multiple Thresholding Binarization Values
to Improve OCR Output presented in Document
Recognition and Retrieval XX Conference 2013,
California, USA, 2013. USA: SPIE.
[12] Shaikh, N. A., & Shaikh, Z. A, 2005, A generalized
thinning algorithm for cursive and non-cursive
language scripts presented in 9th International
Multitopic Conference IEEE INMIC, Pakistan, 2005.
Pakistan: IEEE
[13] Shaikh, N. A., Shaikh, Z. A., & Ali, G, 2008, Segmentation
of Arabic text into characters for recognitionpresented
in International Multi Topic Conference, IMTIC,
Jamshoro, Pakistan, 2008.
[14] Ciresan, D. C., Meier, U., Gambardella, L. M., &
Schmidhuber, J, 2011, Convolutional neural network
committees for handwritten character classification
presented in International Conference on Document
Analysis and Recognition, Beijing, China, 2011. USA:
IEEE.

[15] Archana Balyan, S. S. Agrwal and Amita Dev, Speech
Synthesis: Review, IJERT, ISSN 2278-0181 Vol. 2
(2013) p. 57 – 75.
[16] E. Cambria andB. White, “JumpingNLPcurves:areview
of natural language processing research,” IEEE
Computational Intelligence, vol. 9, no. 2, pp. 48–57,
2014.
[17] van Santen, Jan P. H.; Sproat, Richard W.; Olive,JosephP.;
Hirschberg, Julia (1997). Progress in Speech Synthesis.
Springer. ISBN 978-0-387-94701-3.
[18] Liu, F., Weng, F., Jiang, X.: A Broad-Coverage
Normalization System for Social Media Language. In:
Proceedings of the 50th Annual Meeting of the
AssociationforComputationalLinguistics:LongPapers-
Volume 1. Number July (2012) 1035–1044.
[19] Mark Tathamand Katherine Morton, Developments in
Speech Synthesis (John Wiley & Sons,Ltd. ISBN:0-470-
85538-X, 2005)
[20] “Deep learning for computer vision with Caffe and
cuDNN” NIVIDIA Developer Blog. October 16,2014
[21] (ocr) “A Complete Workflow for Development of
Bangla OCR” International Journal of Computer
Applications (0975 – 8887) Volume 21– No.9, May
2011

Text and Object Recognition using Deep Learning for Visually Impaired People

More Related Content

What's hot (18)

Similar to Text and Object Recognition using Deep Learning for Visually Impaired People (20)

More from ijtsrd (20)

Recently uploaded (20)

Text and Object Recognition using Deep Learning for Visually Impaired People