See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/334130284

An Efficient Scheme for Automatic Pill Recognition Using Neural Networks

Article · June 2019

CITATIONS READS

0 1,526

4 authors:

Muhammad Rizwan Chughtai Gulistan Raja

National University of Sciences and Technology University of Engineering and Technology, Taxila
1 PUBLICATION   0 CITATIONS    70 PUBLICATIONS   445 CITATIONS   

SEE PROFILE SEE PROFILE

Junaid Mir Furqan Shaukat

University of Engineering and Technology, Taxila University of Engineering and Technology, Taxila
25 PUBLICATIONS   65 CITATIONS    18 PUBLICATIONS   143 CITATIONS   

SEE PROFILE SEE PROFILE

Some of the authors of this publication are also working on these related projects:

An efficient scheme for lung nodule detection View project

Digital Techniques View project

All content following this page was uploaded by Furqan Shaukat on 01 July 2019.

The user has requested enhancement of the downloaded file.

 e Nucleus
Th The Nucleus 56, No. 1 (2019) 42-48
The Nucleus
ISS N 002 9-56 98 ( Pri nt)
www.thenucleuspak.org.pk
ISSN 2306-6539 (Online)
Pa
ki sta n

An Efficient Scheme for Automatic Pill Recognition Using Neural Networks

R. Chughtai1*, G. Raja2, J. Mir2 and F. Shaukat3
1
National Institute of Electronics, Islamabad, Pakistan
2
Electrical Engineering Department, University of Engineering & Technology, Taxila, Pakistan
3
Electronics Engineering Department, University of Engineering & Technology, Taxila, Sub Campus Chakwal, Pakistan

ARTICLE INFO ABSTRACT

Article history: An efficient scheme, capable of extracting key pill features, for an automatic pill recognition is proposed.
Received : 01 March, 2019 The devised system involves a number of processes which starts with the thresholding applied to the
Accepted : 25 June, 2019 input query pill image for extraction of the shape feature vector and generation of mask images. The
Published : 27 June, 2019 extracted shape feature vector is used for shape recognition through a trained neural network.
Information regarding the color and size of the pill is obtained by using the mask images and shape
Keywords : information. For pill imprint extraction, a modified stroke width transform (MSWT) and two-step
sampling is applied. The extracted pill query features are compared with the feature values of the created
Imprint recognition,
database for recognition of the pill and its purpose. The proposed method is evaluated on a dataset of
Modified stroke width transform (MSWT), 2500 images and achieves an accuracy of 98% which shows the supremacy of the proposed method in
Neural networks comparison to the other similar pill recognition systems.

1. Introduction recognition system. When the database has a large number of

The pharmaceutical industry is making progress day by pills, the manual input method becomes time-consuming and
day and more effective medicines are being manufactured also requires a human resource to do the inputs. Therefore, for
now-a-days to cure the diseases. With rapid industrial growth, a large number of pills, the automatic pill recognition system
many pharmaceutical industries have emerged and they all are is preferred.
manufacturing their medicines in their unique way. However, 1.2 Automatic Recognition System
with the variety of tablets/pills available in the market, it
An automatic pill recognition system is fast and easy to
raises a problem for an average person to distinguish or
use for a large number of pills as there is no requirement of
recognize an unlabeled pill. While pills can be distinguished
manually entering the data. A lot of research is being carried
based on their appearance (physical appearance
out in the area of automatic pill recognition. Recently, Yu and
characteristics), it demands an effective system which can be
Chen [7] proposed a technique which utilized color, shape and
used to identify/recognize the pills accurately. The Food and
imprint features of the pill for pill recognition and achieved
Drug Administration (FDA), a federal agency in the US
an accuracy of 97.1 %. They created their own pills image
dealing with products related to the food and drugs, enforces
database of 2500 pills. Data augmentation was done by
all pharmaceutical companies through their regulation code
randomly changing brightness, contrast, rotation, etc., of each
21CFR206 [1] to make a unique look of every pill in the
acquired pill image; thereby, making an image dataset of
context of four features. These features are; shape, size, color
12,500 pills images in total. Neto et al. [8] proposed a pill
and imprint of the pill. As every pill has its unique appearance,
feature extractor to classify pills based on used pill shape and
a system can be devised to extract all these four features
color. Feature extractor is evaluated using KNN, SVM and
accurately to predict the unlabeled pill. Several systems have
Bayes classifiers. To extract the features, they utilized two
been developed in recent years to classify and identify an
datasets (PILL BR and NIH NLM PIR) and achieved an
unlabeled pill accurately. These systems can be classified
accuracy of 99.85 and 99.82 in 0.01006 and 0.00810 seconds,
mainly into two main categories; Manual Recognition System
respectively. Wang et al. [9] introduced highlighted deep
and Automatic Recognition System.
learning (HDL) technique for identification of the pill's blister
1.1 Manual Recognition System package. Segmentation and descriptive features can be
The manual recognition system requires a user to provide extracted using HDL technique. The technique utilizes CNN
manual inputs regarding the features of the pill, e.g., color, to classify correct blister pack type and is rotation and light
shape, an imprint, etc. There are many websites which offer invariant. They got almost 100% accuracy on the database of
the online pills recognition system. Few of them include 272 types of blister packages.
WebMD Pill Identification Tool [2], Pillbox [3], RxList Pill Palenychka et al. [10] proposed a new set of rotation-
Identification Tool [4], Drugs.com [5] and Health-line Pill invariant image descriptor, which fully discriminates between
Identifier [6]. Although, the manual input method is easy to different medication pills. They used three independent
utilize, it is less efficient in comparison to the automatic descriptors to form pill descriptor vector. Also, they applied a

Corresponding author : rizwanchughtai1@gmail.com

42
 R. Chughtai et al. / The Nucleus 56, No. 1 (2019) 42-48

machine learning algorithm to train their database for 2116 real drug pill images. Belongie et al. [21] introduced the
detection. Their proposed system shows a precision and recall weighted shape context technique, in which, the similarity
rate of 0.95 and 0.97, respectively. Wong et al. [11] used Deep between two shapes is measured and used for object
Convolutional Network (DCN) for automatic pill recognition.
identification and verification. After the application of
In this paper, an efficient scheme for automatic pill
Geometric transformation, Saliency-driven pill detection is
recognition system using Neural Networks (NN) is proposed.
performed, followed by Deep model training to develop DCN.
In our method, an imprint of the pill is detected by first
They used a database of 400 commonly used tablets and
applying the Modified Stroke Width Transform (MSWT) [22]
capsules. The reported mean accuracy rate of DCN at Top-1
which is a combination of switch function, accumulated
return was 95.35%. Hart [12] proposed object segmentation
gradient and Stroke Width Transform (SWT). Then, Two-step
based mobile computer vision system. Estimators can deliver
Sampling Descriptor Set (TSDS) is utilized to refine the
the parameters (color, shape and size of the pill) of the
imprint section. Thresholding technique is applied for
unknown input pill image, which are used in recognizing the
extracting information related to the shape and size of the pill.
input image by matching these parameters with the database.
Color information is extracted by using a pill mask image and
Nguyen and Huynh [13] worked on tablet imprint recognition.
averaging three color channels of the image. At the end, a
They used polar transform technique and NN to train their
neural network is used for shape and imprint recognition.
system for specific imprint and then recognized the tablet
using a trained network. Hu [14] employed edge localization 2. Proposed Methodology
method and the invariant moment technique for imprint The proposed method consists of a series of steps which
extraction. Chen and Kamata [15] technique is based on start with the analysis of pill image dataset (available on the
finding the measurement of similarity between two shapes. US National library of medicine website [3]). Along with the
The correspondence between two different shape points is pill images, the information about the disease for which the
found and then an aligning transform is evaluated using the pill is prescribed is also available. Information of four pill
correspondence. The technique shows promising results on features, i.e., size, shape, color and imprint, is extracted
different variety of images including handwritten digits, through our devised technique by exploiting the prior
trademarks, silhouettes and the coil data, etc. knowledge about the format of the available pill images in the
Caban et al. [16] introduced a classification technique dataset. A database of acquired features of each pill is created
which automatically identifies the prescription of the tablet in the MATLAB. After the database creation, automatic
using the pill’s image. They used a modified shape feature extraction of any input query pill image (acquired
distribution method and created an invariant descriptor for according to the format and constraints implied in the pill
imprint extraction. Grigorescu and Petkov [17] presented a image dataset acquisition) is performed by our proposed
novel rich local descriptor named distance set which can be system and then compared with the database for matching and
found by using an image feature point. Two-dimensional identifying the pill. Comparison is done by utilizing distance
images are used as a distance set where the distance set metric which computes the distance between the extracted
contains the feature points of the image. By doing so, they feature value of the query pill image and already available
resolved object segmentation and shape matching problem by feature values of the pills present in the dataset. If the distance
finding the difference between distance set and dissimilarity is within the threshold value, the search region for the query
between sets of distance sets. The use of shape image in the database will be reduced. The same procedure is
matching/filtering is very useful in traffic sign detection and performed for all four feature values. Finally, the proposed
recognition of handwritten character. Suntronsuk and system gives the resultant pill image and conveys the
Ratanotayanon [18] presented an algorithm for an automatic information regarding the class of query image along with its
pill recognition system based on the text of the imprint of the location in the database.
pill. Edge-based connected component technique and edge The information about the class and instant (at which the
masking technique is utilized to extract the text from the resultant pill image is present in the database) is utilized to
imprint. Then, an optical recognition system (OCR) is applied find out the disease for which query pill can be used.
to different extracted binary imprint area to recognize the pill. However, if the distance between the query pill image features
Osada et al. [19] illustrated a method to discriminate 3D value and the database pill image features value is not within
shapes. They find a shape measuring function based on the the threshold, the database search range will not be reduced.
signature of an object as a sampled shape distribution, which This will lead to the result that the pill is not present in the
defines the global geometric properties of the object. The database. In our proposed pill recognition system, 2500 pills
discrimination method eventually was used as a pre-classifier images are used as a database. Overview of the proposed
for a recognition or similarity retrieval system. Experimental system is shown in Fig. 1 and detailed description of the steps
results show a 60.9% system accuracy. Lee et al. [20] involved in the devised technique is presented next.
proposed an automatic pill recognition system based on an
imprint of the pill. Feature vector based on edge localization
and invariant moments of the tablet was extracted.
Experimental results showed 93.02% matching accuracy over
43
 R. Chughtai et al. / The Nucleus 56, No. 1 (2019) 42-48

needs to be done. For this purpose, color query pill image is

first converted into the grayscale representation. This is
carried out through appropriate thresholding by converting
each color channel of the input query pill image into the
binary form and then taking a union of all three binary
channels. The threshold value of 0.67 is found to be sufficient
for this purpose for all the pills present in the dataset due to
the uniform background. After thresholding, the resulting
query pill image will have a pill’s shape/geometry represented
in the white region (logical 1) and background represented in
the black region (logical 0) as shown in Fig. 3.

Fig. 3: Binary thresholded image.

Although the shape of the pill is highlighted in the binary

Fig. 1: Block diagram of the proposed pill recognition system. thresholded image, the image needs to be cropped from the
center to separate front and back side of the pill image to
2.1 Shape Identification facilitate the training of the recognition system. The prior
The pill image dataset utilized in our proposed pill knowledge about the pill image dataset is exploited for
recognition system is available on the website of the US separating the front and back side of the pill. Through close
National library of medicine [3]. In this dataset, every pill examination of the dataset, it can be seen that the search
image comprises of the front and back side of the pill region for cropping the pill images lies in between 37% to
presented side by side against a uniform grey background, as 67% of the image resolution as depicted in Fig. 4.
shown in Fig. 2. For pill shape identification, extraction and
recognition of the pill shape from the input query pill image

Fig. 4: Search region for cropping pill image.

Fig. 2: Example of a pill image present in the dataset [3].

44
 R. Chughtai et al. / The Nucleus 56, No. 1 (2019) 42-48

The location for cropping is identified by moving

horizontally a vertical line (equal to the height of the image)
pixel by pixel in the search region and then finding the number
of connected components at each location. For vertically
presented side by side pill images, the same process is
followed with a virtual horizontal line (equal to the width of
the image) moved in the vertical direction. Pixel locations,
where connected component value is 1, indicates that there is
no boundary (shape) present in that location. All such pixel
locations are identified and an average of these pixel locations Fig. 6: (a) Side by side position, (b) Up and down position [3].
is computed to find a suitable location for cropping the binary
image without cutting the front or back side of the query pill. 2.3 Color Identification
The cropped front and back side binary query pill images are To find out the color of the input query pill image, mask
utilized as mask images in color identification process. The images obtained from the shape identification step are
shape of the query pill is recognized through a trained neural utilized. After cropping the input query pill image in a similar
network. fashion as done before, the binary mask image is multiplied
with the colored front and back side of the pill image. As the
2.2 Size Identification
value of the mask is 1 for the pill shape area and zero
The size information of the pill images available on the otherwise, the resulting values after multiplication will give
US National library of medicine website is present within the the pill color values of each R, G and B color channel. The
image in the form of a graph scale (in mm) as shown in average of each color channel is computed to give three color
Fig. 5. The need is to devise a mechanism to extract this values of each channel to be used for color identification
information from the query pill mage automatically. purpose.
2.4 Imprint Identification
To detect and identify imprint on the pill, pre-processing
is first applied to enhance the contrast of the pill image. Then,
the modified stroke width transform (MSWT) [22] is applied
on each color channel separately to get a rough area where pill
imprint exists. Result of application of MSWT on one channel
is shown in Fig. 7.

Fig. 5: Size information of the pill [3].

In order to find the size of the query pill, a binary

thresholded image is first generated to represent graph scales
in white color (logical 1) and the rest of the image in black
(logical 0). As graph scales are already present in the white
color in the dataset pill images, a high threshold value of 0.90
is found to do the needful. The thresholded binary image is
then utilized to find the height and width of the pill by
Fig. 7: Pill imprint after the application of MSWT.
calculating connected components. For this purpose, a
vertical line equal to the height of the image is swept from left Maximally Stable Extremal Region (MSER) algorithm
to right. When the value of the connected component is other [23] is applied to all these three color channel images by
than 1, it indicates the overlapping of the virtual line and setting its region area and threshold equal to 0.1 to get the
vertical graph scales. In this case, the value of the connected maximum output region, as shown in Fig. 8.
components gives the height of the pill under consideration.
The same process is repeated for calculating the width of the
pill in which virtual horizontal line is swept from the bottom
to the top. Finally, the computed width or height is divided by
two based on the orientation of the query pills (as shown in
Fig. 6) in the image to get the size information of the single
pill.
The size information of the pill is stored in the form of
area. The computed height and width information along with
the shape information extracted in the shape identification
stage is utilized to compute the area of the query pill. Fig. 8: Imprint results after the application of MSER algorithm.

45
 R. Chughtai et al. / The Nucleus 56, No. 1 (2019) 42-48

In parallel, the thresholding method is also performed on

the front and back side cropped image of the query pill. A
suitable threshold is estimated as an average of all pixel values
present in the image. Then, the image is converted into binary
form. While converting into binary, the background is
intentionally kept equal to logical 0 (black) and the imprint is
generally mapped to logical 1 (white) as shown in Fig. 9.

Fig. 11 Nine shapes of pills (a: Circular, b: Square, c: Rectangular,

d: Capsule, e: Tear, f: Elliptical, g: Star, h: Diamond, i: Triangular)
[3].

Table 1: Neural network parameters and settings.

Input Images 2500

Fig. 9: Imprint results after the application of thresholding.
Input Layers Equal to Input Sample Size
To get the final imprint image, a union of the image after Hidden Layers 10
thresholding and MSER algorithm is taken to get more Output Layers Equal to Output Values
reliable imprint image as shown in Fig. 10. Activation Function Logsin
Training Function Traingdx
Cost Function MSE
Iterations / Epochs 8000

NN for pill shape recognition in the proposed system is

created with an input layer having nodes equal to the
dimension of the shape feature vector. While the output layer
has nine nodes with each representing nine shape classes,
intermediate hidden layer nodes are equal to 10. For creating
an input shape feature vector, the binary thresholded pill
Fig. 10: Final extracted imprint result.
image from the shape recognition stage is first down-sampled
The dataset of the pills that is used in this research has to 512 x 512 resolution. Then, a column shape feature vector
broad pills diversity in terms of color, shape, size and imprint. of 2601 x 1 dimension is created which contains average
Also, the contrast and brightness of the pills in the dataset values of every 10 pixels present in the row and column of the
have a large variation, so the imprint algorithm gives the image. Finally, the feature vector is normalized to get the
intermediate results. Some imprint extraction results are very binary values. The activation function for all layers is
accurate like the above-shown example and some results are “logsin”, and MSE is used as a cost function. These
just satisfactory. Therefore, a NN is employed to identify the parameters are presented in Table 1. The trained network is
extracted pill imprint correctly. used to identify the shape of the input query pill image by
classifying it into one of the 9 classes. Same parameters and a
3. Experiments and Results similar procedure is also used for imprint recognition.
The proposed method is evaluated on a dataset of 2500 pill However, as each pill imprint is unique, the very large number
images [3]. After extracting all the features, Neural Network of required output layer nodes due to the pill image dataset
(NN) is used for shape and imprint recognition. As there are size makes the usage of NN very complex. To cater this
in total of nine different shapes of the pills, as shown in problem, set of all images with similar shape pills like circle,
Fig. 11, the database is categorized into nine different classes. square, capsule, etc., is created. Then, for each set (which now
These nine shape classes are as follows: contains less number of pill images), NN is trained on the
corresponding final extracted imprint images obtained from
a. Circular b. Square c. Rectangular the imprint identification stage. The selection of trained NN
for imprint extraction of input query pill image is made based
d. Capsule e. Tear f. Elliptical on the shape recognized in the earlier process.
Once the classifier is trained, then during the test phase, a
g. Star h. Diamond i. Triangular query pill image (as a test image, as shown in Fig. 12a) is
given as an input to the proposed pill recognition system. The
system first extracts the features as explained in the
46
 R. Chughtai et al. / The Nucleus 56, No. 1 (2019) 42-48

methodology section and then tends to match these extracted 𝑁𝑢𝑚𝑏𝑒𝑟 𝑜𝑓 𝑓𝑎𝑙𝑠𝑒 𝑟𝑒𝑗𝑒𝑐𝑡𝑖𝑜𝑛 2
𝐹𝑅𝑅 = = = 2% (1)
𝑇𝑜𝑡𝑎𝑙 𝑖𝑑𝑒𝑛𝑡𝑖𝑓𝑖𝑐𝑎𝑡𝑖𝑜𝑛 𝑎𝑡𝑡𝑒𝑚𝑝𝑡𝑠 100
features value with the created pill database in MATLAB. If
a match is found, a resultant matched image is displayed as
shown in Fig. 12b. The system will also present the class and For false images (which does not exist in the dataset), the
instance of the input image in the form of text message such proposed system shows 98 true rejections and 2 false
as “The pill is present in DB. Class =3 and Instance =5. The acceptances.
pill is used for cold and fever”. On the other hand, if the input 𝑁𝑢𝑚𝑏𝑒𝑟 𝑜𝑓 𝑓𝑎𝑙𝑠𝑒 𝑎𝑐𝑐𝑒𝑝𝑡𝑎𝑛𝑐𝑒 2
𝐹𝐴𝑅 = = = 2% (2)
query pill image is not present in the database or no match of 𝑇𝑜𝑡𝑎𝑙 𝑖𝑑𝑒𝑛𝑡𝑖𝑓𝑖𝑐𝑎𝑡𝑖𝑜𝑛 𝑎𝑡𝑡𝑒𝑚𝑝𝑡𝑠 100
the input query image is found, a text message “The pill is not Table 2: Accuracy comparison with other existing systems.
present in the image” will be shown.
Image Image Pixel Size Total No.
Method Accuracy
Database Complexity of Images of Images

Osada Available
High - 133 60.9%
et al.[19] online

Lee et al. Available

Low 1550x2088 2116 93.02%
[20] online

Belongie Available
Medium 416x448 1400 94.2%
et al. [21] online
Fig. 12: (a) Test image [3], (b) Resulting matched image.
Yu and Own
Low 200x200 2500 97.1%
Chen [7] Database
The accuracy of the system can be checked by computing
FRR (False Rejection Rate) and FAR (False Acceptance Proposed Available
High 768x1024 2500 98%
Method online
Rate). To compute the FAR and FRR, a threshold must be
defined first. During retrieval, when a pill query image is
given as an input for recognition, then our proposed method Total error of proposed pill recognition system is:
checks the query image with the existing database and gives 𝑇𝑜𝑡𝑎𝑙 𝐸𝑟𝑟𝑜𝑟 = 𝐹𝑅𝑅 + 𝐹𝐴𝑅 = 4% (3)
the output result in the form of probability, i.e., how much
percentage the query image is matched with the existing
Where, total error is 4% for 200 tested images. The accuracy
database. For example, if an image is matched with 90% of the proposed system is calculated as:
accuracy and a threshold value for the system is equal to 85%,
𝑇𝑜𝑡𝑎𝑙 𝐸𝑟𝑟𝑜𝑟
then the system will recognize it. On the other hand, if an 𝐴𝑐𝑐𝑢𝑟𝑎𝑐𝑦 = 100% −
𝑇𝑜𝑡𝑎𝑙 𝑁𝑢𝑚𝑏𝑒𝑟 𝑜𝑓 𝑎𝑡𝑡𝑒𝑚𝑝𝑡𝑠
image match is 80%, the system will discard the image and
will not recognize it (this falls in the FRR category). If the 4
= 100% − = 98% (4)
threshold value is set high, then FRR will increase and if the 200

threshold is set too low then FAR will increase. The threshold Pill recognition results of our system in comparison with
of the proposed system is set to EER (Equal Error Rate) by hit other systems are shown in Table 2. It can be observed that
and trial method. EER is the point where FAR is equal to FRR our proposed system and devised techniques for feature
as shown in Fig. 13 [24]. extraction are more accurate and reliable. This can be seen as
our system has the highest accuracy of 98 % in comparison to
the other similar methods. Further, our pill recognition system
is also more accurate than Yu and Chen method [7] who have
utilized the same dataset as ours.
4. Conclusions
An automatic pill recognition system is proposed which
can be used in the pharmaceutical factories as a quality
inspection where mass production of the pills is happening.
Four pill features are extracted automatically and then
Fig. 13: FAR-FRR Rate [24].
compared with the database for classification and recognition
of the pill. The proposed system achieves high accuracy of
To check the accuracy of the proposed pill recognition 98% showing its potential for marking unlabeled pills. In
system, 200 query pill images are tested. Of these tested future work, the pill recognition system will be tested on a
images, 100 images exist in the dataset referred to as true very large pill dataset. Further, an effort will be made to build
images. The rest 100 images are false images, which do not the proposed system as an online system by developing its
exist in the dataset. The proposed system gives 98 true App, so that the user can interact with the system easily.
acceptances and 2 false rejections for true images.
47
 R. Chughtai et al. / The Nucleus 56, No. 1 (2019) 42-48

References [13] T.H. Huynh and T.S. Nguyen, “A new imprinted tablet recognition
algorithm using polar transform and neural networks”, Int. Conf. on
[1] CFR- Code of Federal Regulations Title 21, “US food and drug
Advanced Technologies for Communications (ATC), pp. 38-43, 2015.
administration homepage”, Available: http:// www.accessdata.fda.
gov/scripts/cdrh/cfdocs/cfcfr/CFRSearch.cfm?CFRPart=211, 2015 [14] M.K. Hu, “Visual pattern recognition by moment invariants”, IRE
Trans. Inf. Theory, pp. 179-187, 1962.
[2] WebMD Pill Identification Tool, “Pill identification homepage”,
Available: http://www.webmd.com/pill-identification/ [Accessed: [15] Z. Chen and S.I. Kamata, “A new accurate pill recognition system using
2018]. imprint information”, Sixth Int. Conf. Machine Vision, London, UK,
pp. 906711, Dec. 2013.
[3] National Library of Medicine, National Institutes of Health, United
States, “Pillbox homepage”, Available: http://pillbox.nlm.nih.gov/ [16] J.J. Caban, A. Rosebrock and T.S. Yoo, “Automatic identification of
developer.html [Accessed: 2018]. prescription drugs using shape distribution models”, IEEE 19th Int.
Conf. Image Process., pp. 1005-1008, 2012.
[4] RxList Pill Identification Tool, “Pill-identification-tool homepage”,
available at http://www.rxlist.com/pill-identification-tool/article.htm [17] C. Grigorescu and N. Petkov, “Distance sets for shape filters and shape
[Accessed: 2018]. recognition”, IEEE Trans. Image Process., vol. 12, no. 10, pp. 1274-
1286, 2003.
[5] Pill Identification Tool, “Pill identification homepage”, Available:
https://www.drugs.com/pill_identification.html [Accessed: 2018] [18] S. Siroratt and R. Sukanya, “Pill image binarization for detecting text
imprints”, 13th Int. Conf. Computer Science and Software Engineering
[6] Healthline Pill Identifier, “Pill-identifier homepage”, Available:
(CSSE), pp. 978-983, 2016.
http://www.healthline.com/pill-identifier [Accessed: 2018].
[19] R. Osada, T. Funkhouser, B. Chazelle and D. Dobkin, “Shape
[7] J. Yu and Z. Chen, “Accurate system for automatic pill recognition
distributions”, ACM Trans. Graphics, vol. 4, pp. 807–832, 2002
using imprint information”, IET Image Process., vol. 9, no. 12,
pp. 1039-1047, 2015. [20] Y.B. Lee, U. Park and A.K. Jain, “Pill-ID: Matching and retrieval of
drug pill imprint images”, IEEE Proc. 20th Int. Conf. Pattern
[8] M.A.V. Neto, J.W.M. de Souza, P.P. Reboucas Filho and W.D.O.
Recognition, pp. 2632-2635, August, 2010.
Antonio, "CoforDes: An invariant feature extractor for the drug pill
identification", IEEE 31st Int. Symp. on Computer-Based Medical [21] S. Belongie, J. Malik and J. Puzicha, “Shape matching and object
Systems (CBMS), Karlstad, pp. 30-35, 2018. recognition using shape contexts”, IEEE Trans. on Pattern Analysis and
Machine Intelligence, vol. 24, no. 4, pp. 509-522, 2002.
[9] J.S. Wang, A. Ambikapathi, Y. Han, S.L. Chung, H.W. Ting and C.F.
Chen, "Highlighted deep learning based identification of [22] B. Epshtein, E. Ofek and Y. Wexler, “Detecting text in natural scenes
pharmaceutical blister packages”, IEEE 23rd Int. Conf. on Emerging with stroke width transform”, IEEE Computer Society Conf. on
Technologies and Factory Automation (ETFA), Turin, pp. 638-645, Computer Vision and Pattern Recognition , pp. 2963-2970, 2010.
2018. [23] H. Chen, S.S. Tsai, G. Schroth, D.M. Chen, R. Grzeszczuk and
[10] R. Palenychka, A. Lakhssassi and M. Palenychka, "Verification of B. Girod, “Robust text detection in natural images with edge-enhanced
medication dispensing using the attentive computer vision approach", maximally stable extremal regions”, IEEE 18th Int. Conf. on Image
IEEE Intl. Symp. on Circuits and Systems (ISCAS), Florence, pp. 1-4, Process., pp. 2609-2612, 2011.
2018. [24] M.A. Rahman, R.A. Kabir, Z. Begum and M.M. Haque, "A study of
[11] Y.F. Wong, H.T. Ng, K.Y. Leung, K.Y. Chan, S.Y. Chan and C.C. human recognition using inner joining lines of fingers,"5th Int. Conf.
Loy,” Development of fine-grained pill identification algorithm using on Computer Sciences and Convergence Information Technology,
deep convolutional network”, Journal of Biomedical Informatics, Seoul, pp. 186-191, 2010.
vol. 74, pp. 130-136, 2017.
[12] A. Hartl, “Computer-vision based pharmaceutical pill recognition on
mobile phones”, Proc. 14th Central European Seminar on Computer
Graphics, pp. 5, 2010.

48

View publication stats