IAES International Journal of Artificial Intelligence (IJ-AI)

Vol. 12, No. 4, December 2023, pp. 1873∼1882

ISSN: 2252-8938, DOI: 10.11591/ijai.v12.i4.pp1873-1882 r 1873

Biologically inspired deep residual networks

Prathibha Varghese, Arockia Selva Saroja
Department of Electronics and Communication, Noorul Islam Centre for Higher Education, Tamil Nadu, India

Article Info ABSTRACT

Article history: Many difficult computer vision issues have been effectively tackled by deep
neural networks. Not only that but it was discovered that traditional residual
Received Sep 10, 2022
neural networks (ResNet) captures features with high generalizability, render-
Revised Dec 10, 2022 ing it a cutting-edge convolutional neural network (CNN). The images classified
Accepted Dec 21, 2022 by the authors of this research introduce a deep residual neural network that
is biologically inspired introduces hexagonal convolutions along the skip con-
Keywords: nection. With the competitive training techniques, the effectiveness of several
ResNet variations using square and hexagonal convolution is assessed. Using
Convolutional neural networks the hex-convolution on skip connection, we designed a family of ResNet ar-
Deep learning chitecture,hexagonal residual neural network (HexResNet), which achieves the
Hexagonal pixel highest testing accuracy of 94.02%, and 55.71% on Canadian Institute For Ad-
Residual vanced Research (CIFAR)-10 and TinyImageNet, respectively. We demonstrate
Square pixel that the suggested method improves vanilla ResNet architectures’ baseline im-
age classification accuracy on the CIFAR-10 dataset, and a similar effect was
seen on the TinyImageNet dataset. For Tiny- ImageNet and CIFAR-10, we saw
an average increase in accuracy of 1.46% and 0.48% in the baseline Top-1 accu-
racy, respectively. The generalized performance of advancements was reported
for the suggested bioinspired deep residual networks. This represents an area
that might be explored more extensively in the future to enhance all the discrim-
inative power of image classification systems.

This is an open access article under the CC BY-SA license.

Corresponding Author:
Prathibha Varghese
Department of Electronics and Communication Engineering, Noorul Islam Centre for Higher Education
Kumaracoil, Thuckalay-629180, Tamil Nadu, India
Email: prathibha@sngce.ac.in

1. INTRODUCTION
Deep convolutional neural networks have greatly advanced computer vision [1]-[4]. Networks with
a greater number of nodes are better able to capture the subtleties of high-dimensional visual data that lack
linearity. The following factors, however, cause network performance to decline as network depth increases:
– Vanishing gradient: gradients (partial derivatives) are calculated with the aid of the chain rule in backprop-
agation. Throughout this training, gradients typically decrease at an exponential rate optimized using the
activation function being used, meaning that the network’s gradient gets less and smaller [5], [6].
– Harder optimization: it has been discovered that an increase in the number of layers in neural networks
corresponds to an increase in training errors [7].
The problem of vanishing gradients is directly addressed by the design of such a residual network
(ResNet) He et al. [1] which includes several skip connections in addition to the basic convolution layers. Con-
trary to conventional convolution layers, these connections make it simple for gradients to propagate backward

Journal homepage: http://ijai.iaescore.com

1874 r ISSN: 2252-8938

without degrading. Additionally, ResNet is frequently employed as the primary feature extractor in numer-
ous computer vision applications due to skipping connections’ inherent benefits and its generalization capacity
[8]-[10].
This paper presents hexagonal residual networks (Hex-ResNet), a hybrid design with biological inspi-
ration that enhances deep residual network generalization with hexagonal convolutional filters. We demonstrate
that adding small hexagonal filters along a few skip links can improve the ResNet architecture’s base perfor-
mance. The main strength of our strategy is how well it combines the benefits that each of the separate (square
and hexagonal) tessellations has to offer. We implement Hoogeboom et al. [10] which proposes effective pro-
cedures for hexagonal convolution by utilizing an comparable total of two square convolutions, in contrast to
Steppa and Holch [11] which operates directly on a hexagonal lattice. By applying adequate computational sup-
port based on square tessellations, we can increase performance in this way. In comparison to several existing
ResNet models, our suggested design has increased testing and validation accuracy based on Top-1 and Top-5
accuracy. Also, demonstrated how these improved concerts were achieved lacking appreciable computational
rise. We provide the following summary of our efforts and findings:
– We improved the baseline picture classification accuracy of the vanilla ResNet by adding hex convolutions
with a couple of skip connections.
– By conducting extensive tests on the benchmark datasets CIFAR-10 and TinyImageNet, we verified the
effectiveness of the recently suggested Hex-ResNet architecture.
– We demonstrate that, across various ResNet settings, improving the accuracy of image categorization from
scratch by adding hex convolutions to the skipped connection paths.
The rest of this paper is organized as: section 2 discusses all the studies that are linked to our method-
ology, and a survey of the associated literature is provided. Section 3 first covers the fundamentals of traditional
ResNet topologies, skip connections, and then our suggested Hex-ResNet architecture in depth. We provide the
experimental findings and training methods for our suggested architecture in section 4 utilizing the CIFAR-10
and TinyImageNet datasets. In section 6, we draw a conclusion based on our observations.

2. RELATED WORKS
2.1. Residual networks
Deep residual networks are a perfect network to serve as the backbone feature extractor for different
computer vision tasks since they can circumvent the vanishing gradient problem using skip connections [1],
[9], [12], [13]. Li and He [14] presented a convex k-method employing various area parameter altering criteria
and offered an enhanced ResNet via changeable shortcut connections Wightman et al. [15] for numerous
ResNet configurations in the Timm open-source toolbox, pre-trained models and shared competitive training
parameters were made available. The study of Schlosser et al. [16] added pre-activation ResNets by rearranging
the building block’s components to enhance the signal propagation path. All of the well-known efforts that have
ResNet as their primary feature extractor are best suited for data that is defined on a square lattice [15], [16].
2.2. Hexagonal convolution operations
Let S be the equivalent image described on square lattice tessellations, and H be the input image data
representation on hexagonal lattice tessellations. We denote the hexagonal kernels by Kl , where l indicates
its size. We have assumed that the kernel weights are one for mathematical demonstration simplicity. But we
employ the trainable kernel weights in the final implementation. As shown in Figure 1, we generate compa-
rable rectangular kernels (K1r1 ∈ R2×3 and K1r2 ∈ R3×1 ) corresponding to K1 . Remember that a size one
hexagonal kernel will have two similar rectangular shaped kernels. Similar to this a hexagonal kernel of size
l will have l + 1 comparable rectangular kernels. Convolutions with these kernels K1r1 and K1r2 can be now
simply done with efficient PyTorch routines. However, we must suitably pad S in three distinct ways to produce
hexagonal convolutions using rectangular kernels, We must properly pad Sin three different ways as shown in
Figure 1. Let S1 , S2 , S3 be each of the three padded variations of S. When rectangular kernels are convolu-
tioned mathematically K1r1 and K1r2 with S1 , S2 , S3 can be designed as:

P1 = S1 ∗(1,2) K1r1 (1)

P2 = S2 ∗(1,2) K1r1 (2)

Int J Artif Intell, Vol. 12, No. 4, December 2023: 1873–1882

Int J Artif Intell ISSN: 2252-8938 r 1875

P3 = S3 ∗(1,1) K1r2 (3)

Convolve input with S1

kernel of size 1
0 0 0 0 0

1 9 K1 0 1 5 9 13
5 13 Merge
2 10 1 0 2 6 10 14
6 14 1 1
11 * 1 0 3 7 11 15
3 P12
15 1 1 Q
7
4 12 1 0 4 8 12 16 5 22 18 19 8 33 37 46
8 16
H 11 26 38 63 17 44 68 63
Padding 1, kernel 1, stride (1,2)
S2 13 30 42 68 Add 22 31 75 68

1 5 9 13 0 15 16 46 43 22 31 69 43

2 6 10 14 0
1
3 7 11 15 0 Receive output
Kr1 of equal
1 5 9 13 dimension
4 8 12 16 0
1 1
2 6 10 14
0 0 0 0 0
1 1
3 7 11 15
Padding 2,kernel 1,stride(1,2) S3
4 8 12 16 P3 8 37
0 0 0 0 33 46
3 11 19 27
S 17 68
1 Padding 3,kernel 1 5 9 13
2,stride(1,1) 6 18 30 42 44 63
1 22 75
2 6 10 14
9 21 33 45 31 68
1 3 7 11 15 22 69
7 15 23 31 31 43
1
Kr2 4 8 12 16

Figure 1. Implementation of hexagonal convolution [13] for hexagonal lattice image

where P1 , P2 , P3 indicate the convolution outcomes with the kernels K1r1 and K1r2 . The convolution operator
∗(x,y) denotes with stride of x and y units along the rows and columns, respectively. The following step is to
integrate P1 and P2 by selecting the alternate columns as shown in Figure 1. Mathematically we represent the
merge operation as:

P12 = M ERGE (P1 , P2 ) (4)

The square equivalent of hexagonal convolution is obtained by one final addition operation as:

Q = P12 ⊕ P3 (5)

where ⊕ denotes the element-wise addition operation. The output Q if more processing is required, is reorga-
nized into a hexagonal lattice as shown in Figure 1.

2.3. Hex-ResNet
The proposed Hex-ResNet is shown in Figure 2. The skip connections used projection shortcuts
with hexagonal convolution and trained as mentioned by He et al. [1]. The 34-layer Hex-ResNet is with the
integration of square convolution and hexagonal convolution is analysed. Different variants of Hex-ResNet is
developed similar to ResNet 34 architecture.

Biologically inspired deep residual networks (Prathibha Varghese)

1876 r ISSN: 2252-8938

Figure 2. Proposed Hex-residual network architecture (readers are requested to zoom in to view the details)

Int J Artif Intell, Vol. 12, No. 4, December 2023: 1873–1882

Int J Artif Intell ISSN: 2252-8938 r 1877

3. EXPERIMENTAL RESULTS
In this section, We go into detail about our CIFAR-10 experiment outcomes. We also employ the
TinyImageNet dataset for showing the efficacy of our method. Following He et al. [1], We use Top-1 and
Top-5 accuracy as the performance metrics.

3.1. Software and hardware

In order to compute hexagonal solutions, we use the PyTorch-based tool HexagDly [11]. The input
image hexagonal framework is preserved over its output side thanks to a combination of precise padding and a
striding strategy. It uses Google Colab Pro+, a platform for web applications. This contains several integrated
packages that let us train our model on powerful GPUs. The majority of the Tesla V100-SXM2-16 GB and
A100-SXM4-40GB GPUs with compute capacities of 7.0” and 8.0” respectively were used throughout our
studies.

3.2. Datasets
3.2.1. CIFAR-10
The CIFAR-10 benchmark dataset is the industry standard for classifying images. The proposed
architecture is tested on this dataset of 600,000 images organized into 10 categories. The test set is made up of
10,000 images, while the training set is made up of an initial 50,000 photos. Every training image was given
a 4-pixel padding on all sides, and a 32-pixel crop was then randomly selected from either the padding image
or even the padding image’s horizontal flip. The dataset used in the test was not expanded. Both the train
and validation subsets of the augmented train dataset, totaling 5,000 pictures, were created Figure 3 displays
samples of both training in Figure 3(a) and evaluation images from several categories in the CIFAR-10 dataset
in Figure 3(b). The optimum solution was discovered using stochastic gradient descent with the following
parameters: learning rate=0.1, momentum=0.90, and weight decay=0.001. The optimization criterion utilized
was cross-entropy loss. After that, we took the divided learning rate and trained the model for 182 epochs with
32k, 48k, as well as 64k iterations, respectively.

(a) (b)
Figure 3. Sample images from CIFAR-10: (a) training dataset and (b) testing dataset

3.2.2. TinyImageNet
The tinyImageNet dataset, which contains 10,000 training pictures as well as 1,000 validation pictures,
was used to train this Hex-ResNet architecture. Each of its 200 classes contains 500 training images that are
each 64×64 pixels in size Figure 4 displays TinyImageNet’s initial training in Figure 4(a) and testing images
in Figure 4(b). Each subset was then classified as two sets: a train set as well as a validation set, each with an
80:20 ratio. The datasets were enhanced using a variety of approaches, including a) center crop and padding;
b) rotation; c) scaling; d) shearing; e) translation; f) horizontal flip; and g) vertical flip. Stochastic gradient
descent is the optimizer that is employed.

Biologically inspired deep residual networks (Prathibha Varghese)

1878 r ISSN: 2252-8938

(a) (b)
Figure 4. Sample images from ImageNet 2012: (a) training dataset and (b) testing dataset

3.3. Quantitative analysis

The quantitative analysis using the CIFAR-10 and Tiny-ImageNet datasets is depicted in this sec-
tion. Table 1 shows the Top-1 and Top-5 accuracy and error percentage for CIFAR-10 for various ResNet
configurations. The performance of our approach against the baseline residual architectures on CIFAR-10
dataset [17]–[19] is analysed. As observed, our method outperforms all others in terms of validation accuracy,
Top-1 and Top-5 accuracy, and error rates except for Hex-ResNet 32. (Hex-ResNet 20, 44, 56). In addition,
Hex-ResNet 32, 44, as well as 56 exhibits superior validation accuracy with reduced Top-1 error, accordingly,
whenever contrasted to their ResNet equivalents that share the same quantity of layers. Additionally, in con-
trast to ResNet model variant 20 and higher ResNet model 56, Hex-ResNet reduces the Top-1 error by 0.9%
and 0.2%, respectively. Figures 5 (a) to (d) depicts the validation loss vs epochs for both the base ResNet and
Hex-ResNet configurations. Note that Hex-ResNet performs better than ResNet in terms of convergence speed
and validation loss. This clearly shows that when compared to their respective ResNet equivalents, the feature
representation created by the proposed architecture has superior generalization ability and faster convergence.

Table 1. Error rates and accuracy percentage on CIFAR-10. The best scores are indicated by using bold font.
(H) indicates the HexResNet configuration
Validation Top1 Top 1 Top 5 Top 5 Testing
Model Parameters
Acc % acc % error % acc % error % accuracy %
ResNet - 20 272474 91. 92% 91. 44% 8. 56% 99. 63% 0. 37% 91.64%
ResNet - 20(H) 287130 92. 12% 92. 36% 7. 64% 99. 67% 0. 33% 92.12%
ResNet - 32 466906 92. 24% 92. 03% 7. 97% 99. 74% 0. 26% 92. 55%
ResNet - 32(H) 481114 92. 54% 92. 65% 7. 35% 99. 83% 0. 17% 93.14%
ResNet - 44 661338 91. 64% 92. 14% 7. 86% 99. 76% 0. 24% 92.83%
ResNet - 44(H) 675098 92. 92% 92. 94% 7. 06% 99. 92% 0. 08% 93.27%
ResNet - 56 855770 92. 97% 92. 16% 7. 84% 99. 79% 0. 21% 93.07%
ResNet - 56(H) 869082 93. 14% 93. 25% 6. 75% 99. 96% 0. 04% 94.02%

We present the quantitative analysis correspond to TinyImageNet dataset [20]–[22] in Table 2. We

report only the validation scores here due to the unavailability of test set annotations for this dataset. The
results in Table 2 show that proposed Hex-ResNet configurations has lesser validation error compared to their
respective ResNet models. Additionally, all Hex-ResNet variants outperform traditional ResNets and have a
reduced error rate. This demonstrates that the suggested Hex-ResNet architecture is more effectively addressing
the vanishing gradient problem and has a high accuracy percentage. Second, the validation accuracy of the Hex-
ResNet 20, 32, 44, 56, and 110 layers was higher than that of their traditional ResNet counterparts. One of the
important comparisons that show how effective hex residual learning is on incredibly deep systems is this one.
Additionally, Hex-ResNet brought down the error rate by 1.3% compared to the classical ResNet variants. The
advantages of hex residual learning, particularly for deeper network systems, are evident from this comparison.

Int J Artif Intell, Vol. 12, No. 4, December 2023: 1873–1882

Int J Artif Intell ISSN: 2252-8938 r 1879

Lastly, Figures 6(a) to (d) describes the variation of validation loss with respect to epochs for various ResNet
configurations. As can be observed, the Hex-ResNet architecture converges more quickly than the traditional
ResNet architecture. As a result, Hex-earliest ResNet’s stages of convergence are faster and more precise.
Additionally, Hex-ResNet validation loss variation stability is substantially higher than that of its equivalent
ResNet equivalents.

(a) (b)

(c) (d)
Figure 5. Validation loss variation with respect to epochs for CIFAR-10 dataset. ResNet vs HexResNet:
(a) 20 layers, (b) 32 layers, (c) 44 layers, and (d) 56 layers

Table 2. Error rates and accuracy percentage on TinyImageNet. (H) indicates the HexResNet configuration
Model Parameters Validation Accuracy (%) Error Rate (%)
ResNet - 20 2,84,824 48.05 51.95
ResNet - 20(H) 2,99,480 49.51 50.49
ResNet - 32 4,79,256 52.38 47.62
ResNet - 32(H) 4,93,464 52.73 47.27
ResNet - 44 6,73,688 53.65 46.35
ResNet - 44(H) 6,87,448 54.43 45.57
ResNet - 56 8,68,120 55.01 44.99
ResNet - 56(H) 8,81,432 55.71 44.29

Biologically inspired deep residual networks (Prathibha Varghese)

1880 r ISSN: 2252-8938

(a) (b)

(c) (d)
Figure 6. Validation loss variation with respect to epochs for TinyImageNet dataset. ResNet vs HexResNet:
(a) 20 layers, (b) 32 layers, (c) 44 layers, and (d) 56 layers

Compared to ResNet built on pure square tessellations, our technique is less susceptible to noisy
input. The fact that it is visible from both tables is most significant. Table 2 gives the error rates and parameter
comparision on Tiny ImageNet dataset. Most importantly, it can be seen from both the Table 2 in comparison
to the amount of parameters in their respective ResNet counterparts, the number of extra parameters brought
about by hexagonal convolutions is negligibly small.
Table 3 summarizes the best results obtained on CIFAR-10 dataset and Table 4 summarizes the best
results obtained on TinyImageNet dataset using different deep learning models. Results of all the models
trained on scratch is shown in Table 4, shows that the best model is HexResNet network achieving 44.29%
lowest error percentage.

Table 4. Summary of different state-of art-architecture performance on TinyImageNet dataset

Model # Parameters #Error%
InceptionNet [20] 8.3M 56.9
ResNet [11] 11.28M 53.1
VGG-19 [24] 40.2M 49.78
VGG-16 [25] 36.7M 48.09
HexResNet (proposed) 8.8M 44.29

Int J Artif Intell, Vol. 12, No. 4, December 2023: 1873–1882

Int J Artif Intell ISSN: 2252-8938 r 1881

Table 3. Summary of different state of art architecture performance on CIFAR 10 dataset

Model # Parameters #Error (%)
Maxout [15] >6M 9.38
Network in Network [16] ∼1M 8.81
DenseNet [17] 2.5M 8.39
Highway Network [23, 24] 1.25M 8.80
ResNet [11, 25] 0.46M 7.51
HexResNet(proposed) 0.48M 6.73

4. CONCLUSION
In this research work, we proposed biologically inspired hybrid residual network architecture Hex-
ResNet which combines the advantages offered by both square and hexagonal tessellations. We have shown
that using hexagonal convolutions can help us advancing the performance of baseline ResNet architectures on
both CIFAR-10 and TinyImageNet datasets. From the experimental results, we could show that our approach
has better generalisation ability as well as improved convergence properties over the classical ResNet without
increasing significant computational overhead due to hexagonal convolutions. Extension to other computer
vision applications is a potential future direction of our work.

REFERENCES
[1] K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” in Proceedings of the IEEE Com-
puter Society Conference on Computer Vision and Pattern Recognition, Jun. 2016, vol. 2016-December, pp. 770–778, doi:
10.1109/CVPR.2016.90.
[2] A. Krizhevsky, I. Sutskever, and G. E. Hinton, “ImageNet classification with deep convolutional neural networks,” Advances in
Neural Information Processing Systems, vol. 25, pp. 1097–1105, 2012.
[3] M. D. Zeiler and R. Fergus, “Visualizing and understanding convolutional networks,” in Computer Vision–ECCV 2014: 13th Euro-
pean Conference, Sep. 2014, pp. 818–833, doi: 10.1007/978-3-319-10590-1 53.
[4] M. Z. Alom et al., “The history began from AlexNet: a comprehensive survey on deep learning approaches,” arXiv preprint
arXiv:1803.01164, Mar. 2018, doi: 10.48550/arXiv.1803.01164.
[5] Y. Bengio, P. Simard, and P. Frasconi, “Learning long-term dependencies with gradient descent is difficult,” IEEE Transactions on
Neural Networks, vol. 5, no. 2, pp. 157–166, Mar. 1994, doi: 10.1109/72.279181.
[6] X. Glorot and Y. Bengio, “Understanding the difficulty of training deep feedforward neural networks,” in Journal of Machine
Learning Research, 2010, vol. 9, pp. 249–256.
[7] M. S. Ebrahimi and H. K. Abadi, “Study of residual networks for image recognition,” in Intelligent Computing - Proceedings of the
2021 Computing Conference, 2021, pp. 754–763, doi: 10.1007/978-3-030-80126-7 53.
[8] Z. Wu, C. Shen, and A. van den Hengel, “Wider or deeper: revisiting the resnet model for visual recognition,” Pattern Recognition,
vol. 90, pp. 119–133, Jun. 2019, doi: 10.1016/j.patcog.2019.01.006.
[9] M. Farooq and A. Hafeez, “COVID-ResNet: a deep learning framework for screening of COVID19 from radiographs,” arXiv
preprint arXiv:2003.14395, Mar. 2020, doi: 10.48550/arXiv.2003.14395.
[10] E. Hoogeboom, J. W. T. Peters, T. S. Cohen, and M. Welling, “HexaConv,” arXiv preprint arXiv:1803.02108, Mar. 2018.
[11] C. Steppa and T. L. Holch, “HexagDLy—processing hexagonally sampled data with CNNs in PyTorch,” [SoftwareX, vol. 9, pp.
193–198, Jan. 2019, doi: 10.1016/j.softx.2019.02.010.
[12] K. He, X. Zhang, S. Ren, and J. Sun, “Identity mappings in deep residual networks,” in Computer Vision–ECCV 2016: 14th
European Conference, Oct. 2016, pp. 630–645, doi: 10.1007/978-3-319-46493-0 38.
[13] M. Erdmann, J. Glombitza, and D. Walz, “A deep learning-based reconstruction of cosmic ray-induced air showers,” Astroparticle
Physics, vol. 97, pp. 46–53, Jan. 2018, doi: 10.1016/j.astropartphys.2017.10.006.
[14] B. Li and Y. He, “An improved ResNet based on the adjustable shortcut connections,” IEEE Access, vol. 6, pp. 18967–18974, 2018,
doi: 10.1109/ACCESS.2018.2814605.
[15] R. Wightman, H. Touvron, and H. Jégou, “ResNet strikes back: an improved training procedure in timm,” arXiv preprint
arXiv:2110.00476, Oct. 2021.
[16] T. Schlosser, F. Beuth, and D. Kowerko, “Biologically inspired hexagonal deep learning for hexagonal image generation,”
in Proceedings - International Conference on Image Processing, ICIP, Oct. 2020, vol. 2020-October, pp. 848–852, doi:
10.1109/ICIP40778.2020.9190995.
[17] T. Schlosser, M. Friedrich, and D. Kowerko, “Hexagonal image processing in the context of machine learning: Conception of
a biologically inspired hexagonal deep learning framework,” in Proceedings - 18th IEEE International Conference on Machine
Learning and Applications, ICMLA 2019, Dec. 2019, pp. 1866–1873, doi: 10.1109/ICMLA.2019.00300.
[18] J. Luo, W. Zhang, J. Su, and F. Xiang, “Hexagonal convolutional neural networks for hexagonal grids,” IEEE Access, vol. 7, pp.
142738–142749, 2019, doi: 10.1109/ACCESS.2019.2944766.
[19] A. Krizhevsky and G. Hinton, “Learning multiple layers of features from tiny images,” pp. 1–60, 2009, doi: 10.1.1.222.9220.
[20] E. Huynh, “Vision Transformers in 2022: An update on tiny ImageNet,” arXiv preprint arXiv:2205.10660, May 2022.
[21] H. Benbrahim and A. Behloul, “Fine-tuned Xception for image classification on tiny ImageNet,” in 2021 Proceedings of the In-
ternational Conference on Artificial Intelligence for Cyber Security Systems and Privacy, AI-CSP 2021, Nov. 2021, pp. 1–4, doi:
10.1109/AI-CSP52968.2021.9671150.
[22] Z. Abai and N. Rajmalwar, “DenseNet models for tiny ImageNet classification,” arXiv preprint arXiv:1904.10429, Apr. 2019.

Biologically inspired deep residual networks (Prathibha Varghese)

1882 r ISSN: 2252-8938

[23] H. Jung, M. K. Choi, J. Jung, J. H. Lee, S. Kwon, and W. Y. Jung, “ResNet-based vehicle classification and localization in traffic
surveillance systems,” in IEEE Computer Society Conference on Computer Vision and Pattern Recognition Workshops, Jul. 2017,
vol. 2017-July, pp. 61–67, doi: 10.1109/CVPRW.2017.129.
[24] T. Gong and H. Niu, “An implementation of ResNet on the classification of RGB-D images,” in Benchmarking, Measuring, and Op-
timizing: Second BenchCouncil International Symposium, Bench 2019, vol. 12093 LNCS, USA: Springer International Publishing,
2020, pp. 149–155.
[25] M. B. Nourian and M. R. Aahmadzadeh, “Image de-noising with virtual hexagonal image structure,” in 1st Iranian Conference on
Pattern Recognition and Image Analysis, PRIA 2013, Mar. 2013, pp. 1–5, doi: 10.1109/PRIA.2013.6528440.

BIOGRAPHIES OF AUTHORS

Prathibha Varghese received her B.Tech. Degree in Electronics and Communication

Engineering from M.G University, India, in 2005, and M.Tech. Degree in VLSI and Embedded
systems from CUSAT, India, in 2010. She is pursuing Ph.D. in Deep Neural Networks at Noorul
Islam Centre for Higher Education (NICHE), Tamil Nadu. Her research interests are primarily at
the intersection of deep learning and computer vision applications. Specifically, very keen to build
generalizable deep neural network architecture models from very limited training data. She published
many research articles in both Scopus and Sci-indexed journals. She can be contacted at email:
prathibha@sngce.ac.in.

Dr. G. Arockia Selva Saroja received her B.E degree in Electronics and Communi-
cation Engineering from Manonmaniam Sundaranar University, India, in 1997 and M.E degree in
Communication Systems from Madurai Kamaraj University, India, in 1998 and Ph. D Degree from
Noorul Islam Centre for Higher Education, India, in 2017. She is currently working as an Associate
Professor in the Department of Electronics and Communication Engineering. She has been working
in the institution since 1997 and has 15 years of research experience. Her research includes medical
image processing, computer vision, machine learning and wireless networks. She can be contacted
at email: gassaroja@gmail.com.

Int J Artif Intell, Vol. 12, No. 4, December 2023: 1873–1882