Comparison Of Algorithms For Recognition Of Distorted Wagon
Inventory Numbers
Mukhamadieva Kibriyo
1
1
Tashkent University of Applied Sciences, Gavhar Str. 1, Tashkent 100149, Uzbekistan.
https://doi.org/10.5281/zenodo.10471758
Keywords:
Augmentation, segmentation, combined algorithm, recognition, optimality coefficient.
Abstract:
The relevance of the research is due to the need to develop methods, algorithms, and software tools to improve
the efficiency of semantic segmentation of wagon numbers from the video stream in real-time. Despite the
intensive development of modern methods and algorithms, they often do not provide the required quality of
work and reliability, so today there is a need to improve the quality and speed of the semantic segmentation
of objects in the images. Based on the analysis, we have concluded that the most effective solution for the
semantic segmentation is the convolutional neural network CNN with approximated hyperbolic tangent
FastTanh as an activation function and the optimization algorithm ADAM. A convolutional neural network
model with an original architecture consisting of six layers is developed. Software implementation of the
algorithm is done; it allows us to segment more precisely wagon numbers from the video stream in real-time
and to increase the stability and speed of the algorithm in cases of heavily contaminated, low-contrast, and
non-standard wagon number markers. A comparison of the results of different learning algorithms for the
developed neural network is presented.
.
1 INTRODUCTION
The task of automatic detection, segmentation, and
classification of objects is one of the most interesting tasks
of modern computer vision. If in the tasks of classification,
it is necessary to determine only the type of the depicted
object, in the tasks of detection - to construct a bounding
rectangle (or to determine the coordinates) for all objects of
a given type, in the task of semantic segmentation it is
required not only to detect and classify objects but also to
determine their boundaries. In other words, for each pixel
of the image, it is necessary to determine the class of object
to which it belongs. Thus, the task of semantic
segmentation is the most difficult task of image processing.
[17] The difficulty of processing is complemented by the
high variability of objects within one class and the high
similarity of elements of objects of different classes. Of
particular interest is the possibility of solving the problem
of semantic segmentation on computing systems in real-
time.
This article aims to find the most effective way for semantic
segmentation of images from the viewpoint of a
compromise between speed and accuracy. Speed issues are
extremely important for the application of real-time image
analysis algorithms, in the case we are considering, the
recognition of wagon numbers.
This article discusses the chosen method of the algorithm,
training, and activation of neural networks designed for
license plate segmentation. In the experimental part of the
article, the numerical characteristics describing the results
of the combined algorithms under study are analyzed.
Given the fact that the problem we are considering is not
widely covered in publications, we have decided to
compare the existing and newly developed algorithms in
native conditions, comparing the quality and speed of their
work. The conclusion contains a discussion of the results
and main conclusions.
2. Materials and methods
The main approaches to the semantic segmentation of
images include the combined use of three types of
algorithms: detectors, descriptors, and classifiers, which
determine the basic image parameters, select objects, and
classify them. The basic image parameters can be
brightness, color, texture, corners, and borders of objects in
the image and the like.
Among the most popular and effective algorithms that
include detectors and descriptors are SIFT, SURF, FAST,
MSER, and HOG algorithms [23-28].
The SIFT
(Scale Invariant Feature Transform) algorithm
includes a detector and a descriptor. The SIFT detector is
based on the use of scalable spaces - the set of all possible,
smoothed by a particular filter, versions of the same image.
Using a Gaussian filter, this scalable space becomes
invariant to shifts, rotations, and scaling, which does not
shift local extrema. Three parameters are used to determine
the key points: the displacement from the exact extremum
using the Taylor polynomial; the contrast value of the
difference Gaussian; and finding the point on the object
boundary using the Hesse matrix. Then the orientation of
the key point is calculated based on the direction of
gradients of neighboring points [23, 24].
The SURF
(Speeded Up Robust Features) algorithm is an
upgrade of the SIFT detector, but instead of the Gaussian
function, it uses a rectangular 99 filter to approximate it,
thus speeding up the result of the algorithm. In the SURF
descriptor, a square area is built around the point of interest
and divided into square sectors in which the responses to
the Haar wavelets, directed vertically and horizontally, are
computed. These responses are weighted and summed for
each of the sectors [25].
The FAST
(Features from Accelerated Segment Test)
algorithm does not require the calculation of brightness
derivatives but compares the brightness in a circle from the
tested point. First, a quick test of four points from the tested
one is carried out, and then the others are tested. The
number of tests and their sequence is determined on the
training sample [26].
The MSER
(Maximally Stable Extremal Regions)
algorithm is based on determining the pixel intensity of the
image and comparing it with some threshold (if the pixel
intensity is greater than the threshold, it is considered white,
otherwise - black). Thus, we build a pyramid of images with
white images at the beginning and black images at the end.
Such a pyramid allows one to construct a set of coherent
intensity components that are invariant to affine
transformations [27].
The HOG
(Histogram of Oriented Gradients) algorithm is
a key point descriptor based on counting gradient directions
in local image regions. The image is divided into small
coherent regions, which are called cells, and for each cell,
a histogram of gradient directions and edge directions for
pixels within a cell is calculated. The output of the
descriptor is a combination of these histograms [28].
The advantages of these algorithms include high stability to
various geometric and photometric transformations and
image scaling. The disadvantage of these algorithms is the
low stability of operation when the registration angles,
illumination conditions, and reflective surfaces change.
Especially in cases of heavily contaminated, low contrast,
and non-standard wagon number markers.
Among the classifiers for semantic segmentation, various
variants of CNN are most actively used
Faster-RCNN
[19] introduced the regional suggestion
network (RPN) to replace selective search, which makes
Faster-RCNN faster and gives higher accuracy. However,
the region proposal stage is still a bottleneck due to the use
of the selective search algorithm
FCN
[20] fully convolutional encoder-decoder-based
networks are widely used for dense image labeling tasks.
"Encoder" networks are typically backbone CNNs that use
cascading and convolutional levels to learn semantic
information about objects. In contrast, the "decoder" parts
are usually up sampling or deconvolution operations to
recover the lost spatial resolution of encoded features.
SegNet
[29] has a similar design but uses pooled indexes to
record and recover spatial information.
RefineNet
[11] strengthens the decoder by multilevel
function fusion of different levels. Multilevel function
fusion is further enhanced in Exfuse [6] by using both pixel
sum and concatenation operations. The connection between
high-level and low-level functions is also introduced in
DeepLabv3+ [12]. DenseASPP [13] and UNet++ [14]. One
of the limitations of these decoder-encoder designs is that
there is a significant loss of spatial detail at the encoding
stage, and the decoders are still not powerful enough to
recover all the lost information.
Of interest is the development of a segmentation algorithm
that applies machine learning techniques to analyze a string
image with little or no additional preprocessing or post-
processing (end-to-end). Such approaches are distinguished
by the fact that they do not require manual fine-tuning for
a particular case but require a representative training
sample of sufficiently large size. This makes it possible to
simplify and accelerate the creation of segmentation
algorithms for new types of recognized objects, as well as
to increase the accuracy and robustness of various
distortions arising in the imagery.
A special feature of our approach is the use of the
convolutional neural network CNN with an error back
propagation algorithm, L2-regularization, Mini-batch
gradient descent method and as an activation function
FASTTANH fast hyperbolic tangent approximation and
ADAM optimization algorithm for semantic segmentation
2.1. DEVELOPMENT OF A
CONVOLUTIONAL NEURAL
NETWORK
In recent years, CNNs have shown high results in solving
problems of object classification on images. The efficiency
of this approach is explained by the fact that convolutional
neural networks are flexible tools and allow adapting their
structure and parameters to solve the task at hand.
Most approaches to building semantic segmentation
algorithms involve the following steps:
1. Data preprocessing.
2. Pre-segmentation.
3. Feature description.
Classifier training and classification.
5. Context-aware post-processing.
It can be noted that the algorithms have a modular structure,
which allows for choosing different methods at each stage
and their combination.
To date, there are no clearly regulated rules for the
implementation of CNN structure: the number and
organization of layers, the number, and size of feature
maps, the size of convolution matrices, and the choice of
the learning algorithm. CNN is based on the principles of
local perception and separable weights. Local perception
implies that the input of one neuron does not receive all
outputs of the previous layer, but only a certain part of them
[3, 7].
Convolutional neural networks have a much smaller
number of tunable parameters. Also, this type of neural
network is very robust to scaling, shifting and rotating and
other input data transformations [7-9].
The main goal of the experiments was to build the
configuration of the neural network with the smallest
number of parameters. In the process of experimental
studies, CNN’s of different architectures were
implemented, including different numbers of parameters.
Experiments showed that neural networks with simplified
architecture and a small number of parameters showed the
worst results. By sequentially complicating the CNN
architecture, we managed to find the optimal architecture
which ensured high classification results (Figure 1). Further
experiments on complicating the architecture and
increasing the number of CNN parameters did not improve
the quality of classification, but the network operation and
learning time increased significantly.
Figure 1 The architecture of the developed convolutional
network
The experimental neural network was built using the Caffe
framework [2]. This neural network consists of 6 layers and
includes 3 convolutional layers, 1 subsample layer, and 2
fully connected layers. Color images are used as input data.
The input layer has a size of 64*64 neurons. This layer does
not perform any transformations and is only intended to
feed it with input data.
After the input layer, the first hidden layer C1 is located.
This layer is convolutional and contains 64 feature maps,
each of which has the size of 16*16 neurons. The
convolution matrix has a size of 44 neurons. The
displacement is performed by 4 neurons.
The second hidden layer P1 is a subsampling layer, it
consists of 64 feature maps, each of which has the size of
88 neurons. The convolution matrix has a size of 22
neurons. The shift is performed by 1 neuron. This layer
reduces the size of the previous layer by half.
The third hidden layer C2 is convolutional and consists of
112 feature maps, each of which has the size of 66 neurons.
The convolution matrix has a size of 22 neurons. The
displacement is performed by 1 neuron.
The fourth hidden layer C3 is also convolutional and
consists of 80 feature maps of size 33 neurons. The
convolution matrix has a size of 33 neurons. The
displacement is performed by 1 neuron.
The fifth hidden layer FC1 is fully convolutional. This layer
consists of 4096 neurons and has a structure in the form of
a one-dimensional vector.
The sixth hidden layer FC2 consists of 256 neurons and also
has a structure as a one-dimensional vector.
The first four layers of the network have a two-dimensional
structure and are designed to extract features from the
image. The last two layers have a one-dimensional vector
structure and are designed to classify the features extracted
from the previous layers. At the output, the neural network
generates a vector of 256 values, which is converted into a
two-dimensional matrix of 16*16 pixels in grayscale. The
values of each pixel of the output image range from 0 to
255. Initialization of synaptic coefficients of the network
was set randomly in the range from 0 to 1.
When developing the neural network structure, it is also
necessary to select an activation function that is designed to
calculate the output signal of the artificial neuron.
The hyperbolic tangent function was chosen to solve this
problem because it has several advantages [4], which are as
follows:
•
it is symmetric about the origin and provides faster
convergence compared to the logistic function;
•
it has a simple derivative;
•
it is easily differentiable, which simplifies training of
the network by the backward error propagation method
•
has a maximum of the second derivative at = 1.
The hyperbolic tangent function has a range of values from
-1 to 1. This allows the dynamic range of the sigmoid to be
used twice in training to give negative values to the output
signals in the classification layers. The hyperbolic tangent
is given by the formula [4]:
𝑓 (𝑥) = 𝑎 𝑡𝑎𝑛ℎ(𝑏𝑥 ) = 𝑎
(𝑒
𝑏𝑥
−1)
(𝑒
𝑏𝑥
+1)
;
(1)
where a and b – constants.
Figure 2 Hyperbolic tangent
However, the use of the hyperbolic tangent in the network
with a large number of neurons leads to a slow-down of the
calculation and learning process, this is because it is
required to calculate the exponential function which affects
the CPU time.
To solve this problem the algorithm FASTTANH [5] based
on POSIT arithmetic was developed.
It is known that the sigmoidal function is
𝑠𝑖𝑔𝑚𝑜𝑖𝑑(𝑥) =
1
(𝑒
−𝑥
+ 1)
(2)
then the hyperbolic tangent of
𝑡𝑎𝑛ℎ(𝑥)
can be expressed
as:
𝑡𝑎𝑛ℎ(𝑥) =
(𝑒
2𝑥
−1)
(𝑒
2𝑥
+1)
= 2 ∗ 𝑠𝑖𝑔𝑚𝑜𝑖𝑑 (2 ∗ 𝑥)– 1
(3)
From this formulation, an equivalent is constructed which
uses only L1 operators to construct an approximated
hyperbolic tangent. Since we are dealing with 0 exponent
bits of POSIT, all the computation is just a matter of
manipulating the bits, thus efficiently and quickly
computable [5].
Thus, going from sigmoid to a fast version called
FastSigmoid, the approximated hyperbolic tangent looks
like this:
𝐹𝑎𝑠𝑡𝑇𝑎𝑛ℎ(𝑥) = −(1 − 2 ∗ 𝐹𝑎𝑠𝑡𝑆𝑖𝑔𝑚𝑜𝑖𝑑(2 ∗ 𝑥))
(4)
Figure3 Comparison of approximated and real hyperbolic
tangent
Such approximation of the hyperbolic tangent in contrast to
k-tanh
gives an insignificant loss in accuracy of 0.3%, with
a gain in calculation speed of 1.5-2 times [5].
2.2.1CHOICE OF THE TRAINING
ALGORITHM
Neural network training is the sequential correction of
synaptic weights between neurons. One of the most
common and effective learning algorithms for neural
networks is the error back propagation algorithm [10, 14].
The algorithm gets its name from the fact that the error
calculated at each iteration propagates through the ANN
from the output to the input to reconfigure the synaptic
weights. In the process of training the network, when the
input vector is fed, the network output is compared with the
output from the training sample, forming the error [14]. The
correction of synaptic weights is performed by the
following formula [14]:
𝐸
𝑘
=
1
2
∑
(𝑡
𝑘𝑗
− 𝑥
𝑘𝑗
)
2
𝑁
𝑗=0
(5)
where
𝑡
𝑘𝑗
- learning rate coefficient;
𝑥
𝑘𝑗
- neuron input
value;
Value of the network neuron error is defined by the formula
[14]:
∆𝑤
𝑖𝑗
= −𝜂𝛿
𝑘𝑗
𝑥
𝑘𝑗
I(6)
where
𝜂
-learning rate factor;
𝑥
𝑘𝑗
- neuron input values;
𝛿
𝑘𝑗
-neuron error;
𝛿
𝑗
(𝑞)
= (𝑓
𝑖
(𝑞)
(𝑆))′ ∑ 𝑤
𝑖𝑗
𝛿
𝑗
(𝑞+1)
𝑗
(7)
𝛿
𝑗
(𝑞)
- value of the error of the
i
-th neuron in the layer
q
;
𝛿
𝑗
(𝑞+1)
- value of the error of the
j
-th neuron in the layer
q+1
;
𝑤
𝑖𝑗
- weight of the connection connecting two
neurons;
(𝑓
𝑖
(𝑞)
(𝑆))′
- value of derivative activation function
of the
i
-th neuron in the layer
q
.
To regularize the network, we used L2 regularization,
which is a large penalty for a too high value of the weight,
and a small one for a low value, which is expressed in the
use of the regularization coefficient.
We add to the error function a component that is
proportional to the square of the weight’s values
𝐶 = 𝐸 +
𝜆
2𝑛
∑
𝜔
𝑖
2
𝑛
𝑖=1
(8)
𝜕𝐶
𝜕𝜔
𝑖
=
𝜕𝐸
𝜕𝜔
𝑖
+ 𝜆𝜔
𝑖
(9)
This forces the weights to be small, except when the error
gradient is large
The advantages of this learning algorithm are:
•
ease of implementation,
•
ability to use many loss functions,
•
ability to apply large amounts of data.
The disadvantages of the algorithm include small
correction of weights, which leads to a long learning
process. This raises the problem of selecting the optimal
step size. Too small step size leads to slow convergence of
the algorithm, too large step size can lead to loss of stability
of the learning process [14].
To solve these problems, there are various optimization
methods for this algorithm. Out of many existing
optimization methods for training and subsequent
comparison of their performance, Adam (Adaptive moment
estimation), an optimization algorithm that combines the
principles of momentum accumulation and gradient
frequency conservation, was chosen. This method has
advantages of Nesterov accelerated gradient [1], and
AdaGrad [4]. The algorithm, unlike others, does not fall
into the traps of local minima. Adam optimization can
improve the performance of a wide and deep neural
network [21][22].
Also, during training, the minimization of the loss function
was performed using the Mini-batch gradient descent
method [16].
2.2.2 TRAINING AND TESTING THE
DEVELOPED ALGORITHM
For training and testing the developed CNN, a database of
images consisting of several thousand images of wagons
was used. The size of each image is 1920*1080 pixels.
To improve stability, artificial expansion (augmentation) of
the training sample was used using data transformation.
Synthesis of each sample was carried out by applying a
random
set
of
transformations,
simulating
the
transformation of the real field image.
To
expand
the
training
sample,
the
following
transformations were applied (modeling errors of the
system in real conditions): addition of Gaussian noise
distortion, projective distortion to simulate non-ideal
finding, Gaussian blur to simulate defocusing, image
stretching in height and width, vertical and horizontal shifts
and mirror reflections. The following are illustrations of the
described transformations.
Conversion
Illustration
Original image
Gaussian noise
Projective distortions
Gaussian Blur
Shifts
Reflections
Stretch
Transformations Combination
Figure 4. Augmentation of training samples
All images are grouped into training, test, and validation
samples in the ratio of 0.7/0.2/0.1. As seen in Figure 2, the
images contain different classes of objects. The main
objects of interest for the task at hand are wagon numbers.
Figure 3 shows images of segmented objects. These images
correspond to the original images from the training sample
and are intended for CNN training. In the process, CNN
processes small portions of the input images according to
the size of the input layer (64*64 pixels). Thus, the input
image is sequentially scanned by a window of 64*64*64
pixels in size. At each location of this window, the neural
network performs image feature segmentation, forming a
map of 16*16 pixels at the output. This difference in size is
since when sampling a small portion of the image, it is often
difficult to know what is depicted on it (Figure 4). The
increased size of the input image area allows for saving
some data for more effective classification (Figure 4). To
avoid the problem of overtraining, in the fifth full-link layer
is implemented method DropOut [15], which is that during
the training from the overall network is repeatedly and
randomly allocated to a certain subnet, and update weights
occur only within this subnet. Neurons fall into a
subnetwork with a probability of 0.5.
The following neural network parameters were used in
training and testing:
- 0.0005 learning coefficient;
- frequency of learning coefficient change 104;
- the training coefficient variation value is 0,1;
- attenuation for L2 regularization 0,0005.
The configuration of the network remained unchanged. The
number of training epochs for each case was 400.
3. COMPARISON WITH OTHER
IMPLEMENTATIONS OF
CONVOLUTIONAL NETWORKS
When developing segmentation methods, as in the
development of any algorithms, it is necessary to fix a way
to assess the quality of their performance. This method
should allow for the comparison of the developed method
with other algorithms. Let us describe the quality indicators
used in this paper to evaluate methods of wagon number
segmentation.
The purpose of text segmentation into symbols is its
subsequent recognition, which determines the popularity of
using the final recognition quality as an evaluation of
segmentation algorithm quality. An estimate of the quality
of the recognition algorithm can be both the accuracy of
recognition of individual characters or words, and the
average Levenshtein distance. The indicators of quality of
the wagon number recognition system in this work were the
accuracy of full recognition within a symbol because of the
high cost of a single error in a single field - an error of even
a single digit is critical.
In all experiments as the recognition algorithm was used
Tesseract, with default settings. All experiments were done
with the following hardware: Intel Core i5-6400 processor,
8GB RAM, NVIDIA Quadro K5200 graphics card.
Let's look at examples of images of numbers of wagons
taken by us in the working railway station:
a)
b)
Figure. 5. Groups of Wagon Numbers
The images of the wagon numbers were divided into
several groups
a) Poor quality wagon numbers, with high contamination
and low contrast
d) images having inscriptions close in size to the wagon
numbers, as well as in some cases being placed on the same
line.
In the course of our experiments with the segmentation of
wagon numbers with the above-mentioned algorithms, we
made a selection of 100 images for each group. We
encountered the problem of low segmentation accuracy in
images of groups (a) and (b). The problems of segmentation
in the group (b) were partially compensated by comparing
the coordinates of the resulting segments.
Table. 1-4 shows the results of image segmentation by
convolutional neural networks.
Table 1 shows the results of the experiments:
Segmentation algorithm
a
b
SIFT
15,2
77,8
SURF
35,3
71,5
FAST
26,4
81,7
MSER
12,7
74,9
HOG
47,2
72,3
Given that the frequency distribution of the groups in the
sample of 5000 images was the following ratio:
Table №2. The ratio of parameters of the frequency
distribution of groups in a sample of 5000 images
Parameter
Image group (accuracy %)
a
b
Quantity
2500
600
Frequency
0,5
0,12
The final accuracy of the algorithms is calculated according
to probability theory, and we get the result shown in Table
3
Table №3. The result of the final accuracy of
algorithms calculated by probability theory,
Segmentation algorithm
Total accuracy %
SIFT
50,9
SURF
59,4
FAST
56,4
MSER
46,2
HOG
64,2
As can be seen from Table 3, the result of the existing
algorithms is unsatisfactory in real conditions. Due to this
result, it was decided to develop a segmentation algorithm
based on a different approach.
Table №4. Performance of convolutional network
optimization algorithms
Name of algorithm
Time of
training
Accuracy
%
h
min
Nesterov accelerated gradient
10
46
78,13
AdaGrad
27
77,97
Adam
24
85,31
CONCLUSION
As can be seen from Table 4, our trained convolutional
network provides with the Adam optimization algorithm
the best results relative to the others and high enough
segmentation efficiency. The training time was 10 h 24
min, and the accuracy of classification was 85,31%. Almost
all wagon numbers on the images were accurately singled
out, however, there are errors, mainly since in some areas
of the image numbers have a weak contrast to the rest of the
background and are poorly distinguishable. Thus, in the
future, it is planned to conduct experiments with algorithms
to improve image quality, contrast, and application of
various filters.
REFERENCES
[1] Botev A., Lever G., Barber D. Nesterov’s Accelerated
Gradient and Momentum as approximations to Regularised
Update Des_ cent // Machine Learning. – 2016. – V. 1. – P.
1–7.
[2] Caffe
deep
learning
framework.
URL:
http://caffe.berkeleyvision.org
[3] Spatial Pyramid Pooling in Deep Convolutional Networks
for Visual Recognition / K. He, X. Zhang, S. Ren, J. Sun //
Transactions on Pattern Analysis and Machine Intelligence
(TPAMI). –2015. – V. 4. – P. 534–542.
[4] Olgac, A & Karlik, Bekir. (2011). Performance Analysis of
Various Activation Functions in Generalized MLP
Architectures of Neural Networks. International Journal of
Artificial Intelligence And Expert Systems. 1. 111-122.
[5] Cococcioni, Marco & Rossi, Federico & Ruffaldi, Emanuele
& Saponara, Sergio.. A Fast Approximation of the
Hyperbolic Tangent when Using Posit Numbers and its
Application to Deep Neural Networks. doi 10.1007/978-3-
030-37277-4_25. Lecture Notes in Electrical Engineering,
vol 627. Springer, Cham (2020)
[6] Zhang, Z.; Zhang, H.; Peng, K.; Xue, H.; Sun, D. Exfuse:
Feature synthesis enhancement for semantic segmentation.
In Proceedings of the European Conference on Computer
Vision, Munich, Germany, September 8-14, 2018; pp. 269-
284.
[7] Efficient BackProp / Y. LeCun, L. Bottou, G.B. Orr, K.R.
Muller //Neural Networks: Tricks of the trade. – Berlin:
Springer, 1998. –44 p.
[8] Krizhevsky A., Sutskever I., Hinton G. ImageNet
Classification with Deep Convolutional Neural Networks //
Conference on Neural Information Processing Systems
(NIPS). – Nevada, USA, 2012. – P. 27–35.
[9] Howard A.G. Some Improvements on Deep Convolutional
Neural Network Based Image Classification // International
Conference on Learning Representations (ICLR). – Banff,
Canada, 2014. – V. 10. – № 4. – P. 652–659.
[10] Kecman V., Melki G. Fast online algorithms for Support
Vector Machines // IEEE South East Conference
(SoutheastCon 2016). –Virginia, USA, 2016. – P. 26–31.
[11] Lin, G.; Milan, A.; Shen, K.; Reed, ID RefineNet: Multipath
Refinement Networks for High-Resolution Semantic
Segmentation. Cpr 2017, 1, 5.
[12] Chen, L.K.; Zhu, Yu; Papandreou, G.; Shroff, F.; Adam, G.
DeepLabv3+ Atron Separable Convolution Encoder-
Decoder for Semantic Image Segmentation. Proceedings of
the European Conference on Computer Vision (ECCV),
Munich, Germany, September 8-14, 2018; With. 801-818.
[13] Jan, M.; Yu, K.; Zhang, K.; Lee, Z.; Yang, K. DenseASPP
for Semantic Segmentation in Street Scenes. In Proceedings
of the IEEE Conference on Computer Vision and Pattern
Recognition, Salt Lake City, UT, USA, June 18-22, 2018; pp.
3684-3692.
[14] Zhou, Z.; Siddiqui, M.M.R.; Tajbakhsh, N.; Liang, J.
Unet++: Nested u-net architecture for medical image
segmentation. In deep learning of medical image analysis
and multimodal learning for clinical decision support;
Springer: New York, NY, USA, 2018; pp. 3-11.
[15] Wager S., Wang S., Liang P. S. Dropout training as adaptive
regularization //Advances in neural information processing
systems. – 2013. – Т. 26.
[16] Haji S. H., Abdulazeez A. M. Comparison of optimization
techniques based on gradient descent algorithm: A review
//PalArch's Journal of Archaeology of Egypt/Egyptology. –
2021. – Т. 18. – №. 4. – С. 2715-2743.
[17] Mukhamadieva K.B. Image processing method for
recognizing the desired objects/ Conference: Innovative
ways to solve current problems, 2020
[19] S. Ren, K. He, R. Girshick, and J. Sun, “Faster r-CNN:
Towards real-time object detection with region proposal
networks,” in Neural Information Processing Systems
(NIPS), 2015.
[20] Jonathan, L.; Shelhamer, E.; Durrell, T. Fully Convolutional
Networks for Semantic Segmentation. In Proceedings of the
IEEE Conference on Computer Vision and Pattern
Recognition, Boston, Massachusetts, USA, June 7-12, 2015;
With. 3431-3440.
[21] Mohd Jais I., Ritahani A., Qamrun S. Adam Optimization
Algorithm for Wide and Deep Neural // Network Knowledge
Engineering and Data Science (KEDS) pISSN 2597-4602
Vol 2, No 1, Juni 2019, pp. 41–46
[22] Rasmus B., Rohl´en A. A comparison of training algorithms
when training a Convolutional Neural Network for
classifying road signs // 142X EECS/KTH 2019
[23] Ivashechkin A.P., Vasilenko A.Yu., Goncharov B.D.
Methods of finding the singular points on image and their
descriptors. Young cientist, 2016, no. 15, pp. 138–140. In
Rus.
[24] Park S., Yoo J.H. Realtime face recognition with SIFT based
local feature points for mobile devices. The 1st International
Conference on Artificial Intelligence, Modelling and
Simulation (AIMS 13). Malaysia, 2013. pp. 304–308.
[25] Tawfiq A., Ahmed J. Object detection and recognition by
using enhanced Speeded Up Robust Feature. International
Journal of Computer Science and Network Security, 2016,
vol. 16, no. 4, pp. 66–71.
[26] Tore V., Chawan P.M. FAST Clustering Based Feature
Subset Selection Algorithm for High Dimensional Data.
International Journal of Computer Science and Mobile
Computing, 2016, vol. 5, no. 7, pp. 234–238.
[27] Mammeri A., Boukerche A., Khiari E. MSER based text
detection and communication algorithm for autonomous
vehicles.
IEEE
Symposium
of
Computers
and
Communication. Messina, Italy, 2016. pp. 456–460.
[28] Dalal N., Triggs B. Histograms of Oriented Gradients for
Human Detection. IEEE Computer Society Conference on
Computer Vision and Pattern Recognition (CVPR). San
Diego, USA, 2005. Vol. 1, pp. 886–893.
[29] Badrinarayanan, V.; Kendall, A.; Sipolla, R. SegNet: Deep
Convolutional Encoder-Decoder Architecture for Image
Segmentation. IEEE Trans. Model analog. Car. Intelligence.
2017, 39, 2481–2495. [CrossRef] [PubMed]
