FPGA-based Hardware Acceleration for Fruit Recognition Using SVM

Irish Interdisciplinary Journal of Science & Research (IIJSR)
Vol.6, Iss.2, Pages 22-29, April-June 2022
ISSN: 2582-3981 www.iijsr.com
22
FPGA-based Hardware Acceleration for Fruit Recognition Using SVM
Udhaya Kumar C1
, Miruthula R C2
, Pavithra G3
, Revathi R4
& Suganya M5
1-5
Department of Electronics and Communication Engineering, Sri Eshwar College of Engineering, Coimbatore, India.
DOI: http://doi.org/10.46759/IIJSR.2022.6204
Copyright © 2022 Udhaya Kumar C et al. This is an open access article distributed under the terms of the Creative Commons Attribution License,
which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.
Article Received: 12 February 2022 Article Accepted: 17 April 2022 Article Published: 24 May 2022
1. Introduction
In the 80s, the development of agriculture step by step combined with computing to promote management
automatically [1]. Nowadays, rising methods like the Internet of Things additional create additional
management for the agricultural economy [2]. The farms, for instance, will use a variety of sensors to monitor
the environment of expansion environments of crops and their health statuses.
By mistreatment of the collected detector knowledge, farmers and firms will extract the data to boost the
productivity and production of crops. Moreover, with the AI recognition, the potency management of
agriculture so increased considerably [3],[4].
The targeted crop for recognition may be a valid application in object detection. Currently, the convolutional
neural network is the most representative model of deep learning. A CNN consists of the input layer and
multiple hidden layers associated with an output layer. The hidden layers embrace a series of convolutional
layers with twisted multiplication of scalar products.
Therefore, CNN’s area unit is computing intensity that hat they're sometimes enforced on exhibiting
platforms likewise cloud servers [5]. Furthermore, this requires that the acquired pictures be transferred
through a communication network to those powerful platforms, such as cloud servers, to facilitate object
detection. However, this sometimes ends up in high latencies of information transfers within the net, and so
the period detection cannot be achieved [7].
In upcoming years, as luck would have it, a brand new different known as edge computing which projected to
bring knowledge and computation to the supply once needed. The hardware acceleration system is
ABSTRACT
Selection classification for Fruit recognition could be an absolute zone of inspection. Fruit Recognition mistreatment FPGA-based Hardware
Acceleration by SVM is helpful for the observance and indexing of the fruits consistent with their kind with the peace of mind of a quick
production chain. During this test, we have processed to initial replacement prime quality data-set of pictures grouped in the 5 preferred
varieties of oval-shaped fruits. Honor to the fast image process techniques for the development, image resolution, quality of the algorithms
leads to carry-out image process and computational tasks. In recent years, deep neural networks have a diode to the event of the many new
applications associated with preciseness agriculture, as well as fruit recognition. An algorithm consumes computer power and memory, which
has a significant impact on standard and performance, especially when working with large image datasets. Within the planned work, FPG is
A based mostly on hardware acceleration for fruit, and recognition is mistreatment with SVM. The Support Vector Machine could be a
real-time machine learning tool meant for high predicted classification accuracy through the attributes mentioned. Using SVM for embedded
system programs is incredibly difficult attributable to the intensive computations needed. This will increase the attractiveness of implementing
SVM on hardware platforms for reaching performance computing with the demanded value of power consumption. Finally, a difficult
trade-off between meeting embedded period systems constraints and high classification accuracy has been determined.
Keywords: Fruit recognition, FPGA, Support vector machine, Hardware acceleration.

23
the custom-made answer to enhance programs by mistreatment architectures proposed for multiprocessing.
Field Programmable Gate Array offers hardware design and development to be implemented at a lower price.
The most renowned version of FPGAs includes general-purpose logic resources and registers
with specialized modules like memory, slice, and multipliers. During this research, we tend to approach two
plans like plants and crop management, which can permit finding the fruits out of their leaves [6]. That
system is in absolute demand once it involves a system for feeding fruits and robots that supervise the
execution of vegetables and fruits [8],[9].
Each approach area is supported by the SVM classifier.
2. Basic Principles
A modern supervised learning instance, support vector machines, is considered one of the most effective
machines learning methods, enabling sensible generalization performance for a varied array of regression and
classification tasks [11]. The supervised learning area is comprised of two distant phases, the coaching, and
the classification or it can be regression, just in cases where the system’s output is synchronous. The Support
Vector Machine part is to report for the identification of those points of knowledge which will construct a
separate for the variant. These vector units are then unable to find the variant of future datum throughout the
classification section [Fig.1].
The Support Vector Machine helps to learn, supportive info of extraordinary learning mechanism for
classification and support, it can be then applied online or offline depending on its need.
Fig.1. Fruit Recognition by SVM Classifier using Deep Learning Features
However, the Support Vector Machine division could be a technologically costly phase, with stable growth in
the classification load [12]. If wherever the huge scale of issues in a unit targeted or a high range of classified
outputs should be established, the classification phase becomes wildly time-of-frame substandard and
imperatively performed for acceleration rises. Further, on the targeted demerits, the data-set will be
characterized as heterogeneous.
The Uniform datasets area unit typically resembles image capturing, likewise in face recognition and
detection. Under uniform datasets, the selectness needs among the data-set alternative area of identity [10].
For instance, MNIST consists of twenty-eight × twenty-eight-bit pictures per coaching sample and their
uniform drawback, since every 784 indexed from their samples need an associate degree or angle eight-bit bit

24
under illustration. Still, another real-time data set promotes a vital role in diversities among the vigorous
ranges of its proportions.
The variables among heterogeneous datasets would be categorical, indexing, mathematician, or continuous.
As an example, the attributes of adult datasets embrace, varied among others, a person’s status, these variables
will be characterized as continuous, boolean, and categorical, and completely different exactness needs arise
between them.
If one wants to solely represent a one-bit attribute mathematician, the legal entities, in keeping with the
dataset specifications, enter into seven distinct classes and a three-bit illustration is sufficient. The
performance and development of the SVMs will be maximized by customizing the utilization of the
on-the-market computational resources to enquire under certain considerations about the character of the
input file. This is one of the viable platforms for targeting Support Vector Machine classifiers on computing
devices, which may utilize the potential of custom exactness arithmetic [Fig.2].
Fig.2. Hyperlite of the heterogeneous SVM classifier
FPGAs area unit semiconductor devices, that contain programmable logic components and a hierarchy of
programmable interconnects. These days, FPGAs contain, additionally, coarse grain elements, like embedded
multipliers, digital signal process blocks, and memory blocks.
The implementation of the laborious logic array, like multipliers onto the programmable material has enabled
the FPGA devices to spice up their performance potency [13]. FPGA devices supply an extensive quantity of
parts part and various memory sizes, providing a huge range of workable and huge quantity of parallel
procedure power.
Furthermore, its mobility permits it to be used as a solution wherever the application suffers from information
measure limitations. The FPGA reconfiguration offers a major advantage against application-specific
customary merchandise and integrated circuits respectively, once targeting completely different classification
issues which can vary in dynamic vary constraints, size, and spatial property. To boot, trendy FPGA devices
area units able to supply equal or superior performance at a lower power price than general-purpose graphic
process units.

25
3. Existing Methods
This section, combined selective enforcement in machine learning on embedded systems and strategies in
hardware acceleration. The quickest way to detect edge method is to complete the tasks, thanks to parallel
computational proficiency of FPGA to extend method rate and scale back moderate timing, the bundle rule
with XfOpencv has few merits in process time collate to different settings. A lot lately Tzanos et al. exhibit a
hardware acceleration supported They Lomas Bayes to enhance ARM processors, the projected method will
scale back the compiling time resembles sixteen, eight times whereas coaching half and fourteen times in the
detected part [13]-[15].
Implementation through the learning mechanism is predicated on 3 notations, the initial one is to in-build
much information in BRAM to linear access from the FPGA to cut back transferring information latency, and
the second purpose is to avoid operation bottlenecks by partitioning massive arrays into multiple smaller
individual arrays, the last purpose is to pipe-lining loops that assured the utmost similarity.
Different techniques projected a VLSI design of Thomas Bayes classification in FPGA for on-time
extractions of facial expressions, this approach will perform real work extraction operational at a frequency of
241, 55 MHz [16]-[18].
4. Methodology
A-frame to ease accomplish the program for deep learning algorithms treatment PYNQ workflow is planned,
the answer can facilitate skill man of science and hardware to mix the employment of Deep learning model
with design FPGA design always. Our methodology relies on a group of initializing functions before the
extraction section. the dimensions of the info are 5000 pictures broken into 2 categories (“2500 pictures
contain fruits and 2500 pictures of tree leaves”) to conciliate image size. At the beginning of the feature
extraction section, we tend to re-sized the pictures to possess a common size of 150 x 150 x 3, blend bar graph
of orientating gradient HOG, we tend to withdraw the pictures options to be classified; the final pace consists
in coaching SVMs to possess a focused classification model. The hypothesis regarding the acceleration of the
algorithmic program is to Associate in overlay dedicated to the part of the re-sizing of images pixels and to
judge the compiling period and hence the performances of the feature extractions, while not acceleration.
A. Dataset
The dataset to be examined in the performance of the prompt methodology includes pictures of forty forms of
Indian fruits (Fig.1). These pictures consisted of a “13 Megapixel smartphone camera in natural daylight
while not shade”. The dataset contains twenty-three, 848 numbers of pictures and makes offered [19],[20]
[Fig.3]. The study of online implementation of light-weight with a superior hardware accelerated answer. To
take as a baseline the well-known Support Vector Machine that separates work-flows. The unit delineated by
easy packet-level options, the scale of the first 3 points within the flow [21],[22]. As with any supervised
technique, the SVM formula consists of 2 main parts: a coaching part and a detection phase. Throughout the
part, the formula initiates from classes of applications and computing the featuring model by the separating
fruit. This model denoted part decides of the class of application of recent workflow.

26
Fig.3. ModelSim Environment
B. Support vector machine
Support Vector Machine could be a triple-crown extraction in supervised learning. It shows extended
progress in beholding application, the most merits of the SVM is to seek out the best to increase the divisions
of 2 categories just in case of binary way of feature classification and extraction. Every point of the
categories has 2 alternatives parallel in units created and the algorithms try, and then denote the most
effective way of separating the maximum space pixels.
C. Image resizer
Fig.4. Details of image Established for Fruit Recognition

27
The method towards the standardization size is knowledgeable, since the information pictures square measure
massive. To decrease the debugging time frame, the method has a way to set back the scale of the images
properly. The scale of the information (5000 images) is given, which will take longer to size them to 150 x
150 x 3.
FPGA Board is associated with PYNQ Z1, which is equipped in the framework Pynq and in Zynq7020 and
processor induced in the identical chip. Certainly, the processing system incorporates twin-core ARM cortex
A9 computer hardware equipped with a built-in Ubuntu UNIX system, as well as the programmable logic PL
(FPGA) unit equipped with associate Artix seven family memory components, that includes 13300 logic
slices, 630 K of quick block RAM and 220 DSP slices, 512 MB of DDR3 with 16-bit bus. Pynq complies with
the hardware acceleration requirements, and it includes Python drivers for FPGA bitstream transfer and
information transmission using an application programming interface (API). Overlays would be used for
programmable logic circuits when applied as hardware content.
5. Result and Conclusion
In this study, we experimented the classification models for the best performance with the six powerful
architectures in deep neural networks technique for recognition of different kinds of fruits. The experimental
studies were enforced victimization of the MATLAB-2019 deep learning tool.
Fig.5. Simulation result of SVM classifier
The active performance of each classifier is in terms of Specificity, Sensitivity, Accuracy, alphabetic
character coefficient and, Precision, False Positive Rate (FPR). To perform the fine-tuning supported transfer
for deep learning models from pre-trained CNN networks [Fig.5] [Fig.6].
Fig.6. Simulation result of SVM classifier with Fruit Recognition

28
The testing parameters for transfer learning, like most epochs, small size, initial learning rate, and also the
validation frequency were allotted as 0.001 and 30, 5, 64. Furthermore, the random gradient descent with
momentum was meant as a training methodology. Therefore, the performance of Fruit Recognition using
FPGA-based Hardware Acceleration by SVM is best altogether sense.
Declarations
Source of Funding
This research did not receive any grant from funding agencies in the public or not-for-profit sectors.
Competing Interests Statement
The authors declare no competing financial, professional and personal interests.
Consent for publication
Authors declare that they consented for the publication of this research work.
References
[1] K.Kapach, E. Barnea, R. Mairon, Y. Edan, O. Ben-Shahar, (2012). Computer vision for fruit harvesting
robots. State of the art and challenges ahead, Int. J. Comput. Vis. Robot. 3 4e34, https://doi.org/10.1504/
IJCVR.2012.046419.
[2] Zhang, L. Wu, (2012). Classification of fruits using computer vision and a multiclass support vector
machine, Sensors 12(9): 12489e12505. https://doi:10.3390/s120912489.
[3] T.Thamaraimanalan, S.P.Vivekk, G.Satheeshkumar and P.Saravanan (2018). Smart Garden Monitoring
System Using IOT. Asian Journal of Applied Science and Technology, 2(2): 186-192.
[4] S. Bennedsen, D.L. (2005). Peterson, Performance of a system for apple surface defect identification in
near-infrared images, Biosyst. Eng. 90 419e431.
[5] G. Fernando, A.G. Gabriela, J. Blasco, N. Alexios, J.M. Valiente, (2010). Automatic detection of skin
defects in citrus fruits using a multivariate image analysis approach, Comput. Electron. Agric. 71 189e197.
https://doi:10.1016/j. compag.2010.02.001.
[6] Chaudhari, D., Waghmare, S. (2022). Machine Vision Based Fruit Classification and Grading—A
Review. In: Kumar, A., Mozar, S. (eds) ICCCE 2021. Lecture Notes in Electrical Engineering, vol 828.
Springer, Singapore. https://doi.org/10.1007/978-981-16-7985-8_81.
[7] Rojas-Aranda, J.L., Nunez-Varela, J.I., Cuevas-Tello, J.C., Rangel-Ramirez, G. (2020). Fruit
Classification for Retail Stores Using Deep Learning. In: Figueroa Mora, K., Anzurez Marín, J., Cerda, J.,
Carrasco-Ochoa, J., Martínez-Trinidad, J., Olvera-López, J. (eds) Pattern Recognition. MCPR 2020. Lecture
Notes in Computer Science, vol 12088. Springer, Cham. https://doi.org/10.1007/978-3-030-49076-8_1.
[8] Xue, G., Liu, S. & Ma, Y. (2020). A hybrid deep learning-based fruit classification using attention model
and convolution autoencoder. Complex Intell. Syst. https://doi.org/10.1007/s40747-020-00192-x.

29
[9] J. Lopez, C. Maximo, A. Emanuel, (2011). Computer-based detection and classification of flaws in citrus
fruits, Neural Comput. Appl. 20 975e981, https://doi.org/10.1007/s00521- 010-0396-2.
[10] Fu, Y., Nguyen, M. & Yan, W.Q. (2022). Grading Methods for Fruit Freshness Based on Deep Learning.
SN COMPUT. SCI. 3, 264. https://doi.org/10.1007/s42979-022-01152-7.
[11] A. Krizhevsky, S. Ilya, H.E. Geoffrey, (2012). Imagenet classification with deep convolutional neural
networks. Advances in Neural Information Processing Systems, pp. 1097e1105.
[12] Venkatesan, C., & Karthigaikumar, P. (2018). An efficient noise removal technique using modified error
normalized LMS algorithm. National Academy Science Letters, 41(3): 155-159. https://doi.org/10.1007/
s40009-018-0635-0.
[13] T. Thamaraimanalan and P. Sampath, (2019). Leakage Power Reduction in Deep Submicron VLSI
Circuits using Delay based Power Gating. National Academy Science Letters, 43(3): 229-232.
[14] W. Shi, J. Cao, Q. Zhang, Y. Li, and L. Xu, (2016). Edge computing: Vision and Challenges. IEEE
Internet of Things Journal, 3(5): 637-646.
[15] Felzenszwalb and D. Huttenlocher, (2004). Efficient graph-based image seg- mentation. International
Journal of Computer Vision, 59: 167-181.
[16] H. Meng, K. Appiah, A. Hunter and P. Dickinson, (2011). FPGA implementation of Naive Bayes
classifier for visual object recognition. CVPR 2011 WORKSHOPS, pp.123-128, doi: 10.1109/CVPRW.
2011.5981831.
[17] K. Appiah, H. Meng, A. Hunter and P. Dickinson, (2010). Binary histogram based split/merge object
detection using FPGAs. Proceedings of CVPR 2010 (6th IEEE Workshop on Embedded Computer Vision).
[18] D. J. M. Moss, E. Nurvitadhi, J. Sim, A. Mishra, D. Marr, S. Subhaschan-dra, and P. H. W. Leong,
(2017). High-performance binary neural networks on the Xeon+FPGAT M platform, in Proceedings of the
27th International Conference on Field Programmable Logic and Applications (FPL), pp. 1-4.
[19] J. Qiu, J. Wang, S. Yao, K. Guo, B. Li, E. Zhou, J. Yu, T. Tang, N. Xu, S. Song, Y. Wang, and H. Yang,
(2016). Going deeper with embedded FPGA platform for convolutional neural network, in Proc.
ACM/SIGDA Int. Symp. Field-Program. Gate Arrays, pp. 26-35.
[20] R. Zhao, W. Song, W. Zhang, T. Xing, J.-H. Lin, M. Srivastava, R. Gupta, and Z. Zhang, (2017).
Accelerating binarized convolutional neural networks with software-programmable FPGAs, in Proc.
ACM/SIGDA Int. Symp. Field Program. Gate Arrays, pp. 15-24.
[21] O. Russakovsky, J. Deng, H. Su, J. Krause, S. Satheesh, S. Ma, Z. Huang, A. Karpathy, A. Khosla, M.
Bernstein, A. C. Berg, and L. Fei-Fei, (2015). Ima Genet large scale visual recognition challenge. Int. J.
Comput. Vis., 115(3): 211-252.
[22] W. Li, M. Canini, A. Moore, and R. Bolla, (2009). Efficient Application Identification and the Temporal
and Spatial Stability of Classification Schema, Computer Networks, Special Issue on Traffic classification
and its applications to modern networks, 53: 790-809.

FPGA-based Hardware Acceleration for Fruit Recognition Using SVM

More Related Content

Similar to FPGA-based Hardware Acceleration for Fruit Recognition Using SVM (20)

More from Associate Professor in VSB Coimbatore (20)

Recently uploaded (20)

FPGA-based Hardware Acceleration for Fruit Recognition Using SVM