IAES International Journal of Artificial Intelligence (IJ-AI)

Vol. 12, No. 3, September 2023, pp. 1210~1223

ISSN: 2252-8938, DOI: 10.11591/ijai.v12.i3.pp1210-1223  1210

Image preprocessing and hyperparameter optimization on

pretrained model MobileNetV2 in white blood cell image
classification

Parmonangan R. Togatorop1, Yohanssen Pratama2, Astri Monica Sianturi1, Mega Sari Pasaribu1,
Pebri Sangmajadi Sinaga1
1
Department of Information System, Faculty of Informatics and Electrical Engineering, Institut Teknologi Del, Laguboti, Indonesia
2
Department of Software Engineering Technology, Faculty of Vocational Studies, Institut Teknologi Del, Laguboti, Indonesia

Article Info ABSTRACT

Article history: White blood cells play a role in maintaining the immune system which
consists of several types such as neutrophils, lymphocytes, monocytes,
Received Sep 1, 2021 eosinophils and basophils. MobileNetV2 is one of the pretrained
Revised Sep 10, 2022 convolutional neural network (CNN) models that provides excellent
Accepted Sep 21, 2022 advantages and performance in classifying images. In this research was
conducted to find out how to apply optimization hyperparameters and the
impact of image processing on white blood cell image classification using
Keywords: MobileNetV2, so that it is expected to find a combination of preprocessing
and combination of hyperparameter values that can produce the highest
Convolutional neural network accuracy value. To maximize the classification process, before classifying the
Hyperparameter optimization image, several stages of image preprocessing are carried out, namely
Image classification cropping, grayscale, resizing and augmentation. Hyperparameter tuning was
Image preprocessing carried out for an experiment to improve model performance. The three main
MobileNetV2 parameters used in hyperparameter tuning are learning rate, batch size, and
Transfer learning number of epochs. Performance optimization model performance will be
White blood cell measured using accuracy, sensitivity, specificity and using a confusion
matrix. Based on the experimental results in this study, it shows that the best
learning rate value is 0.00001, the best batch size value is 32, and the best
epoch value is 250.
This is an open access article under the CC BY-SA license.

Corresponding Author:
Yohanssen Pratama
Department of Software Engineering Technology, Faculty of Vocational Studies,
Institut Teknologi Del
Jl. Sisingamangaraja Sitoluama-Laguboti, Toba 22381, Indonesia
Email: [email protected]

1. INTRODUCTION
White blood cells play a role in maintaining human immunity. The types of white blood cells are
neutrophils, lymphocytes, monocytes, and eosinophils. Currently, there have been many studies related to
white blood cell image classification [1]. Each type of white blood cell image has a different shape and
morphology from texture, shape and size. [2]. Yildirim and Çinar [3] classified white blood cell images into 4
types, namely eosinophils, lymphocytes, monocytes, and neutrophils using Alexnet, Resnet50, Densenet201
and GoogleNet. The results of the classification used AlexNet 79.27%, Resnet50 78.74%, DenseNet201
77.88%, and GoogleNet 62.93% [3], [4] classified white blood cell images into 4 types (lymphocytes,
monocytes, eosinophils, and neutrophils) using convolutional neural network (CNN), LeNet and AlexNet. The
classification accuracy results obtained are CNN 75.36%, LeNet 65.21% and AlexNet 80.37% [4].

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

Int J Artif Intell ISSN: 2252-8938  1211

Throngnumchai et al. [2] classified white blood cell images into 4 types, namely Monocytes, Neutrophils,
Eosinophils, and Lymphocytes using MobileNetV2 with an accuracy of 93.18% [2]. Based on several previous
studies, it was shown that the pretrained model of MobileNetV2 has good performance for classifying white
blood cell images.
MobileNetV2 can be considered as an appropriate technique to overcome the problem of image
classification [5]. Dong et al. [5] compared the MobileNetV1, MobileNetV2, and CNN models. Compared to
the three models used, MobileNetV2 has the highest accuracy value of 89.00%. Tobias et al. [6] concluded
that the application of CNN, especially the pretrained model of MobileNetV2 with Linear Bottleneck, can
achieve 90% accuracy [6].
MobileNetV2 has a number of hyperparameters that can produce different classification accuracy for
each of the same tasks with different hyperparameters. Yuningsih and Mustikasari [7] classified anemia based
on red blood cell morphology using the CNN. In this study, experiments were conducted on hyperparameters.
The hyperparameters used for the training algorithm are epoch, mini-batch, learning rate. with an accuracy of
0.9774 [7], [8] conducted a study on the effect of batch size on the performance of CNNs for various datasets.
Xu [9] using hyperparameter learning rate, batch size, and number of epochs. The experimental results show
that when the other hyperparameters remain unchanged, the error trained by the network model decreases
significantly as the learning rate increases, the batch size decreases and the number of epochs accumulates
within a certain range [9], [10] conducted research on the importance of using hyperparameter tuning. Machine
learning algorithms such as neural networks involve a number of hyperparameters that must be defined before
they are executed. Hyperparameters can be determined during training and need to be optimized to achieve
maximum performance. Determination of hyperparameters can minimize generalization errors. Therefore, the
results of image classification accuracy using CNN architecture with pretrained MobileNetV2 model can be
optimized by optimizing hyperparameters.
To support the model's performance in classifying images, there are several factors that influence in
terms of data availability, namely the image focused on the object [11], the amount of data processed is
sufficient [12], the image can be processed with small computations [11], and so on. In a study using medical
images, [11] detected a disease of narrowing of blood vessels in the brain. This study uses several pre-
processing techniques, namely image scaling, image gray-scaling, and image noise removal using several
techniques. In medical image processing, image pre-processing is very important so that the extracted image
does not have noise [2], [11] apply image augmentation and morphology image processing before classifying.
Efficient classification of white blood cell types requires a lot of information for study. This researcher
reproduced the data up to 10,000 images. Novoselnik et al. [12] classifying white blood cell images using the
CNN model where one of the preprocessing techniques used is data augmentation. The available dataset
contains 738 images of blood samples, the dataset is very unbalanced (50% of the dataset is neutrophils) which
can result in a biased model. Therefore, augmentation techniques are used, namely random rotation, horizontal
shift, vertical shift, and horizontal flipping [12]. That results in a higher variance of the training data, which
prevents the model from overfitting. Throngnumchai et al. [2] perform morphology image preprocessing and
data augmentation in white blood cell image classification to improve the quality and increase the number of
white blood cell images. The white blood cell image which initially numbered 410 was augmented to 10,000
data [2]. Based on the literature study that has been described, the researcher will perform several image
preprocessing techniques and hyperparameter optimization to improve the performance of the pretrained
MobileNetV2 model in classifying white blood cell images.

2. RESEARCH METHOD
This section describes the research methods used to classify white blood cell images. The first step is
to download the data to be used from the Kaggle website, that is the white blood cell image dataset. Then
perform several stages of image processing to improve image quality. Transfer learning from pre-trained
MobileNetV2 was carried out and a hyperparameter optimization algorithm was applied to classify white blood
cell image types. Then evaluate the results of the classification of the type of white blood cell image obtained
previously.
The research was conducted using the MobileNetV2 Pretrained Model. MobileNetV2 is a CNN
architecture that works based on an inverted residual structure, where the input and output of the residual block
are bottleneck layers as opposed to the traditional residual model, which uses an extended representation in the
input. The MobileNetV2 architecture consists of 32 fully convolution layers with 32 filters, followed by 19
residual bottleneck layers. The MobileNetV2 model runs faster for the same accuracy across the latency
spectrum. Specifically, the new model uses twice less operation, requires 30% fewer parameters and is about
30-40% faster on Google Pixel phones than the MobileNetV1 model [13]. In MobileNet there is a special layer
known as depthwise separable convolution and could be senn in Figure 1. This layer plays a role in reducing

Image preprocessing and hyperparameter optimization on pretrained model … (Parmonangan R. Togatorop)

1212  ISSN: 2252-8938

complexity and parameters so that it will produce a model with a larger size. In the last stage we will apply
pointwise convolution to reduce the dimensions, the pointwise convolution could be seen in Figure 2.

12 1 8
5
1
1
5
3
3 8
12

Figure 1. Depthwise convolution

1 3
8 8
1

3 1
8 8

Figure 2. Pointwise convolution

One of the developments from the previous version of MobileNet is the Linear Bottleneck. The linear
bottleneck function is to remove the nonlinear activation function from the 1×1 pointwise convolution, to
maintain the features that have been extracted by the convolution, the linear activation function will be used.
The use of linear bottlenecks aims to ensure the expressiveness of the model. To reduce the loss of information,
the output of the dimension layer reduction, namely the bottleneck layer is not connected to the non-linear
activation layer, but is connected to the Linear Bottleneck [14], [15].
MobileNetV2 has been used by several researchers to classify images and the classification results
obtained are good. MobileNetV2 which uses a linear bottleneck can achieve higher accuracy with much less
processing time making it more efficient. Some of the reasons that support researchers using MobileNetV2 are
considered from several advantages [16], [17].
− MobileNetV2 is able to achieve higher accuracy with much less processing time so that it is more efficient
than the CNN model.
− MobileNetV2 has a smaller number of parameters than the original MobileNet so less computational costs
are required.
− MobileNetV2 is a very effective feature extractor for image classification, object detection, and
segmentation.
− MobileNetV2 is an architecture that optimizes memory consumption.
The following Figure 3 is the MobileNetV2 architecture which is composed of several layers. Based
on Figure 3, it can be seen that MobileNetV2 is composed of input layer, functional layer, average pooling,
and dense layer. The input layer will input 224×224×3. The output of the input layer is an array with a size of
224×224×3. The output of the input layer will be processed into the functional layer. The functional layer
consists of several bottlenecks. The number of bottlenecks in MobileNetV2 is 17 blocks. Each bottleneck will
receive and process a different input size. The output of the functional layer is an array with a size of 7×7×1280.
Then the results of the Functional layer will be processed on the next layer, namely GlobalAveragePooling2D.
In GlobalAveragePooling2D will be given a 7×7 filter. The channel given to this layer is None so that the
output size is 1×1×1280. In the Dense Layer, the four types of white blood cells will be classified. Every
bottleneck in MobileNetV2 has its constituent components. The following Table 1 are the components that
make up the bottleneck in MobileNetV2. The research method used to classify white blood cell images can be
seen in the following Figure 4.

Int J Artif Intell, Vol. 12, No. 3, September 2023: 1210-1223

Int J Artif Intell ISSN: 2252-8938  1213

Figure 3. The Architecture of MobileNetV2

Table 1. The components of bottleneck in MobileNetV2

No Bottleneck Components
1 Bottleneck I Depthwise Conv2D
Batch Normalization
ReLU
Conv2D
Batch Normalization
2 Bottleneck II, IV, VII, XIV Conv2D
Batch Normalization
ReLU
Zero Padding
Depthwise Conv2D
Batch Normalization
ReLU
Conv2D
Batch Normalization
3 Bottleneck III, V, VI, VIII, IX, X, XI, XII, XIII, XV, XVI, XVII Conv2D
Batch Normalization
ReLU
Depthwise Conv2D
Batch Normalization
ReLU
Conv2D
Batch Normalization

Figure 4. The general stages of the system in the research

Image preprocessing and hyperparameter optimization on pretrained model … (Parmonangan R. Togatorop)

1214  ISSN: 2252-8938

2.1. Provide white blood cell image dataset

To train and evaluate MobileNetV2, we have utilized the dataset available from the Kaggle website.
The dataset stores 367 images consisting of five types of white blood cells, namely basophils, eosinophils,
lymphocytes, monocytes, and neutrophils. However, in this study, 4 types of white blood cells were used,
namely eosinophils, lymphocytes, monocytes and neutrophils. Available images have been labeled; each
picture contains 1-2 pictures of white blood cells. The white blood cell object in the image file is located in a
non-uniform position. The white blood cell images in the dataset are composed of Red, Green, Blue (RGB)
color modes. The following Table 2 is an image of a white blood cell obtained from Kaggle that will be used
in this study.

Table 2. Example of white blood cell image on the dataset

Image Cell Type Image Cell Type
Eosinophils Monocytes

Lymphocytes Neutrophils

2.2. Image preprocessing

Image pre-processing plays an important role in supporting image processing. This is necessary to
simplify processing because most of the image data is impractical to manipulate. With the implementation of
image preprocessing some irrelevant image features will be reduced to make processing and analysis tasks
more effective and efficient. There are several stages applied in the image preprocessing stage in this study,
namely resizing, cropping, grayscaling, and data augmentation.

2.2.1. Cropping
Cropping is an image cutting process that is used to remove other components that are not included
in the object of study. Cropping is done to focus the main image that will be used in image processing. Before
the image is cropped, the image position (right, left, up, down) will be adjusted according to the desired pixels
to get an image that focuses on the desired main object. After being analyzed there are several images that
contain many other components besides the white blood cell object. The purpose of cropping is to focus on the
main object that you want to extract and eliminate components or noise that are not included in the focus of
the research object [18], [19].
Cropping will be done by using the Numpy and cv2 libraries. At the cropping stage, the bounding box
will be used by utilizing the region of interest (RoI) function. First, the image will be loaded, then define the
RoI function and then determine the bounding box of the image to be cropped. Cropping will be carried out on
all 367 raw white blood cell images by utilizing the RoI function, 367 new white blood cell images will be
generated as well. This cropping process will be repeated 367 times according to the amount of raw data you
want to crop using the same configuration. The following Figure 5 is the result of the cropping process.

Int J Artif Intell, Vol. 12, No. 3, September 2023: 1210-1223

Int J Artif Intell ISSN: 2252-8938  1215

Figure 5. The results of cropping technique on image of white blood cells

2.2.2. Grayscaling
Grayscaling is applied to convert the RGB image color to grayscale. Grayscaling helps to simplify the
RGB layer of the image containing the image information at each layer and obtain the image in grayscale
intensity levels. The application of grayscale on white blood cell images will make computational time
efficient, the storage space required when running the algorithm [11]. The color conversion that will be carried
out in this study will use the formula [20],

I = 0.299 R + 0.587 G + 0.114 B

the color conversion to grayscale will be done by using the cv2 libraries. This grayscale process will be repeated
as many as 367 according to the amount of raw data that you want to grayscale using the same configuration.
The image conversion process begins by loading the previously cropped image as grayscaling input. Then
convert the image color using a formula. The results of the grayscale image will then be saved. The data used
as input in the grayscale process is the image of white blood cells resulting from the cropping process that has
been done previously. The following Figure 6 is the result of the grayscale process.

Figure 6. The results of grayscale technique on image of white blood cells

2.2.3. Data augmentation

The data available on the sources used still amount to 367 files. This amount of data is very less to be
used in the process of implementing image classification so that the augmentation process is needed to produce
more data to get more optimal performance [21]. After the preprocessing of cropping and grayscale is done,
the next step will be the augmentation stage. The augmentation stage carried out in this study will produce
8,000 train data images, each type 2,000, and test data will be generated as many as 2,000 images, each type
500.
Image augmentation in python can be done using the ImageDataGenerator function provided by
Keras. The techniques used for this augmentation are random horizontal flip, random vertical flip, random
image rotation, and zooming. The data used as input in the augmentation process is the image of white blood
cells resulting from the grayscale process that has been done previously. The data will be loaded by defining
the path of the image directory location and then checking the image size. The size of the image is a 3-
dimensional array (224, 224, 3) so it is necessary to change the size of the image dimensions to a 4-dimensional
array by adding 1 dimension, this is because the NumpyArrayIterator must have a 4-dimensional array so that
the result of the image array size becomes (1, 224, 224, 3). The next step is to produce n images. The following
Figure 7 is the result of the image augmentation process where 1 image is augmented into 5 images.

Image preprocessing and hyperparameter optimization on pretrained model … (Parmonangan R. Togatorop)

1216  ISSN: 2252-8938

Figure 7. The results of augmentation technique on image of white blood cells

2.2.4. Resizing
Resizing is used to equalize the size of all images that will be used in image processing. Resizing the
image can speed up the computational process and increase the effectiveness of model training as the image
size gets smaller [18]. The application of image resizing makes the results of feature extraction consistent. In
MobileNetV2, the image size usable on the model is larger than 32×32. However, the larger the image size,
the better the model performance will be. The image size that will be used for this research is 224×224. The
data used as input in the resizing process is the image of white blood cells resulting from the augmentation
process that has been done previously.
Image resizing can be done by using the cv2 and Glob libraries. The program will read the image of
white blood cells resulting from the augmentation process and resize the image according to the input
dimensions using the cv2.resize function. In this study, the image size that will be used is 224×224. The
following Figure 8 is the result of the image resizing.

Figure 8. The results of resize technique on image of white blood cells

2.3. Transfer learning from pretrained model

The use of transfer learning is aimed at improving the performance of the model that will be used in
white blood cell image classification. The use of transfer learning makes it easier for researchers not to carry
out the training process from the beginning. In this study, researchers can use the results of the pre-trained
model to extract features from all white blood cell images in a predetermined dataset [5], [6].
In this study, transfer learning from the Pretrained Model MobileNetV2 will be used to classify white
blood cell images. Researchers will implement the system by loading MobileNetV2 using Keras tools and the
Tensorflow backend on the initial layer, then adding a dense layer to classify white blood cell types. At each
layer on MobileNetV2, the research team used Imagenet as the weight. All weights in the pre-trained model
(MobileNetV2) have been trained with ImageNet to recognize colors, textures, and other patterns. In addition

Int J Artif Intell, Vol. 12, No. 3, September 2023: 1210-1223

Int J Artif Intell ISSN: 2252-8938  1217

to using the pretrained CNN model as a feature extractor and retraining the classifier layer that is fully
connected. Transfer learning can also be done by fine-tuning the weights of the pretrained model [22].
In this study, researchers will do fine-tuning to retrain several convolution layers that have been
determined in number. So that the model will obtain a feature extraction result that is more in line with the data
set because it is obtained from the results of retraining several convolution layers on fine-tuning. The
application of fine-tuning can improve the performance of the model in image classification. In fine-tuning the
model will extract and re-learn the features of the image on the tuned layer to extract specific features that are
more suitable for the dataset so as to improve the model's performance in classifying white blood cell images
according to each type.

2.4. Applying hyperparameter optimization

At this stage the hyperparameter optimization is carried out using the manual hyperparameter tuning
method where the hyperparameter values used will be set manually by combining hyperparameters [23]. The
hyperparameters used in this research are learning rate, batch size and epoch [24]. This research will be
conducted experimentally. Each experiment will collect the results and rank based on the value of accuracy
and loss.
The experiments were carried out sequentially, starting with the learning rate experiment. After
obtaining the best learning rate value, then the batch size experiment was carried out. After obtaining the best
batch size value, the epoch experiment was carried out. The experimental learning rate values are 0.1, 0.01,
0.001, 0.0001, 0.00001, 0.000001, 0.0000001. The values of the batch size that were experimented with were
32, 64, 128, and 256. The epoch values that were experimented with were 50, 100, 150, 200, 250.

2.5. Evaluation of experimental results

After the experiment is carried out, then an evaluation of the model that has been built is then carried
out. In the evaluation process starting from, the data to be processed is the white blood cell dataset. Data first
by dividing the data into train data and test data. Each training and testing process will be evaluated to see how
well the built model classifies the white blood cell image. Evaluation of training and testing is done by
calculating the accuracy value of each model. The test data will be evaluated using the library confusion matrix
to calculate the values of accuracy, sensitivity and specificity [25].

3. RESULTS AND DISCUSSIONS

The results and discussions sections explain the results of the implementation and discuss the results
obtained from the implementation system. In this part, the implementation of the design that have been made
will be explained. The things that will be explained in this part are the implementation environment,
implementation constraints, libraries, image preprocessing implementation, CNN model implementation and
model evaluation and how the results obtained from each classification process carried out in this study.

3.1. Preprocessing result

We have tested the effect of preprocessing on the resulting accuracy value. The test scenario carried
out is to first test the effect of preprocessing on the accuracy value if the preprocessing technique is only
resizing and augmentation. After that, testing the effect of preprocessing on the accuracy value if the
preprocessing technique used is cropping, resizing and augmentation. Based on the Table 3, it can be seen that
the combination of preprocessing cropping, grayscale, augmentation and resizing has a more optimal accuracy
value by considering the resulting test value.

Table 3. Average accuracy results of the preprocessing experiment

Experiment Train Validation Test
Accuracy Loss Accuracy Loss Accuracy Loss
Resize + Augmentation 0.8670 0.3865 0.8302 0.4566 0.4770 1.6311
Cropping + Resize + Augmentation 0.9947 0.0243 0.9906 0.0325 0.8470 0.8803
Grayscale + Resize + Augmentation 0.8662 0.3886 0.8315 0.4517 0.5360 1.5360
Cropping + Grayscale + Resize + Augmentation 0.95774 0.13882 0.93633 0.17748 0.9665 0.0907

In the resize and augmentation experiments, the accuracy value obtained is quite good, but in the test
results, the accuracy value is very low. In the cropping, resizing and augmentation experiments, the accuracy
value obtained is good, but in the test results, the accuracy value is quite good. In the grayscale, resize and
augmentation experiments, the accuracy value obtained is quite good but in the test results, the accuracy value

Image preprocessing and hyperparameter optimization on pretrained model … (Parmonangan R. Togatorop)

1218  ISSN: 2252-8938

is very low. In the cropping, grayscale, resize and augmentation experiments, the accuracy values obtained are
good and in the test results, the accuracy values are also good.

3.2. Hyperparameter optimization

The selection of the appropriate hyperparameter is one of the things that plays a very important role
in CNN training and has a major impact on the performance of the model being built. We conducted several
experiments to select the best hyperparameter values in classifying white blood cell images. The following are
the results and discussion of hyperparameter optimization experiments on learning rate, batch size, and number
of epochs.

3.2.1. Learning rate

Experiments on learning rate are used to determine the learning rate that produces the highest best
accuracy value. The experimental learning rate value is 0.1, 0.01, 0.001, 0.0001, 0.00001, 0.000001,
0.0000001. Based on the Table 4, it shows the learning rate value affects the model’s performance in white
blood classification.

Table 4. Average validation accuracy results of the learning rate experiment

Experiment Learning Batch Number of Average Value of Average Value of
Rate Size Epochs Validation Accuracy Validation Loss
1 0.1 32 60 0.69611 3.26628
2 0.01 32 60 0.91362 1.79781
3 0.001 32 60 0.97009 0.41778
4 0.0001 32 60 0.99037 0.03675
5 0.00001 32 60 0.99098 0.02607
6 0.000001 32 60 0.95049 0.15984
7 0.0000001 32 60 0.53752 1.07380

From the experiments, the learning rate with the highest accuracy was obtained at the application of
learning rate 0.00001 (1×10-5) with an accuracy 0.99098 and a loss value 0.02607. The following Figure 9 is a
graph of the average accuracy and loss validation values for the learning rate experiments. Based on the graph,
it can be seen that in the learning rate range of 0.1 - 0.00001, the accuracy value tends to be higher and the loss
value is getting smaller. This is due to a decrease in the value of the learning rate from 0.1 to 0.00001 which
results in a better model recognizing data, so that the resulting accuracy increases and the loss decreases.

Figure 9. Graph of average validation accuracy and loss from the learning rate experiment

Int J Artif Intell, Vol. 12, No. 3, September 2023: 1210-1223

Int J Artif Intell ISSN: 2252-8938  1219

The model training uses a learning rate of 0.001; 0.0001, 0.00001 produces high accuracy and small
loss. This is because the use of a small learning rate value makes the model speed better at recognizing data.
However, the use of a small learning rate will require a lot of time in conducting training. Meanwhile, model
training using learning rates of 0.1 and 0.01 resulted in poor accuracy and significant loss. This is because the
use of a large learning rate value makes the model speed faster in recognizing data. The use of a large learning
rate tends to require a short time in conducting training. However, based on experiments on the learning rate,
after training the model uses a smaller learning rate value of 0.000001 and 0.0000001, the accuracy value
decreases (smaller) and the loss value increases. Therefore, it can be seen that the model has a good
performance in conducting training not depending on the size of the learning rate but when the learning rate is
at the optimal level.

3.2.2. Batch size

In the batch size experiment, 60 epochs and a learning rate of 0.00001 were used, which is the learning
rate value with the best accuracy in the previous learning rate experiment. The following are the results
and discussion of the batch size experiment carried out for the values of 32, 64, 128 and 256. Based on the
Table 5, it is known that the batch size value affects the model's performance in classifying white blood cell
images.

Table 5. Average validation accuracy results of the batch size experiment

Experiment Batch Learning Number of Average Value of Average Value of
Size rate Epochs Validation Accuracy Validation Loss
1 32 0.00001 60 0.99097 0.02607
2 64 0.00001 60 0.98251 0.05420
3 128 0.00001 60 0.96100 0.13669
4 256 0.00001 60 0.90955 0.31664

The smaller the batch size, the more optimal the accuracy value and the smaller loss value. From the
experiments that have been carried out, it is known that the most optimal batch size=32 value is with an
accuracy value of 0.99097and a loss value of 0.02607. The following Figure 10 is a graph of the average
accuracy and loss validation values for the batch size experiments that have been carried out. Based on the
experimental batch size chart in Figure 10, it can be seen that the smaller the batch size, the higher the accuracy
value and the smaller the loss value and vice versa.

Figure 10. Graph of average validation accuracy and loss from the batch size experiment

Image preprocessing and hyperparameter optimization on pretrained model … (Parmonangan R. Togatorop)

1220  ISSN: 2252-8938

This is because the batch size experiment was carried out using a small learning rate of 0.00001. The
use of a learning rate value of 0.00001 which is included in the small category, will require a small batch size
value as well in order to maintain the stability of the model. Therefore, the value of batch size 32 will be more
optimal than other batch sizes for the use of a learning rate of 0.00001 which is included in the small category.
Theoretically, batch size affects training time but does not significantly affect test performance, therefore the
accuracy value in each experiment does not change significantly.

3.2.3. Number of epoch

To determine the number of epochs that produce the highest best accuracy value, the researcher
conducted experiments on several epochs, namely 50, 100, 150, 200, 250. In this experiment, the learning rate
and batch size used were selected based on the learning rate experiment and batch experiment. size with a
learning rate of 0.00001 and a batch size of 32. Based on the experiments that have been carried out, the
researchers obtained the average value of the validation accuracy from the number of epoch experiments which
can be seen in the following Table 6.

Table 6. Results of the number of epoch experiment

Experiment Epoch Learning Rate Batch size Average Value of Validation Accuracy Average Value of Validation Loss
1 50 0.00001 32 0.98719 0.01547
2 100 0.00001 32 0.99525 0.00844
3 150 0.00001 32 0.99749 0.00695
4 200 0.00001 32 0.99786 0.00488
5 250 0.00001 32 0.99892 0.00380

After conducting several experiments on the number of epochs, namely 50; 100; 150; 200; 250, the
researcher obtained the highest average validation accuracy value in the experiment with a number of epochs
of 250. From the graph in Figure 11, it can be seen that the larger the epoch, the higher the accuracy value and
the smaller the loss value and vice versa. This statement is proven in the experimental results, namely at epoch
50, the accuracy value is 0.98719 and the loss is 0.01547. As the epoch increases, the accuracy increases and
the loss decreases. So that the epoch 250 obtained an accuracy value of 0.99892 and a loss value of 0.00380.
From the experimental results, it can be seen that the number of epochs does not have a significant effect on
the accuracy and loss values. This is because the epoch hyperparameter represents the number of iterations the
learning algorithm will work through the entire training dataset. Therefore, the more iterations given, the better
the accuracy value.

Figure 11. Graph of average validation accuracy and loss from the number of epoch experiment

Int J Artif Intell, Vol. 12, No. 3, September 2023: 1210-1223

Int J Artif Intell ISSN: 2252-8938  1221

4. EXPERIMENTAL EVALUATION
After the experiment is done, then testing is carried out on the model that has the best accuracy value.
The test results will be evaluated using a confusion matrix. The test data used consisted of 2,000 image data,
namely 500 images of each type of white blood cell tested.
The evaluation results obtained in each learning rate experiment using a batch size of 32, number of
epochs 60 and a learning rate value of 0.01-0.000001 resulted in a relatively good test evaluation (testing)
indicated by a good accuracy value, which is close to 1. In this study, the best average model is obtained from
the highest average value in all iterations of the fold. Test results to ensure that the model used is good for
recognizing images with different data sets.
The results of the evaluation carried out on each batch size experiment, it can be seen that for the
learning level of 0.00001 and epoch 60, the batch size of 32-256 produces a relatively good accuracy value on
the evaluation of the test (test). In this study, determining the best model is done by finding the average value
of accuracy in all fold iterations. The test results are used to ensure that the model used performs well to
recognize images with different data sets.
From the epoch experiment, it is known that for a learning rate of 0.00001, batch size of 64, and
epochs 60-250, the evaluation of testing is relatively good which shows the accuracy is close to 1. by finding
the average value across the fold iterations. The test results are used to ensure that the model used performs
well to recognize images with different data sets.

5. CONCLUSION
MobileNetV2 can be optimized by using hyperparameter optimization manually to classify white
blood cell images by changing each hyperparameter value for each experiment performed. Based on the
experimental results, it shows that the best learning rate value is 0.00001, the best batch size value is 32, and
the best epoch value is 250. The use of a relatively small learning rate results in slower model speed in learning,
but a smaller learning rate allows the model to recognize the data better. The use of a smaller learning rate will
require a longer learning time. When the learning rate is higher, the speed of the model increases when it
recognizes data because a large learning rate requires a faster time to train. The smaller the batch size value,
the higher the accuracy value. A high accuracy value will be directly proportional to the smaller loss value and
vice versa. Batch size does not really show a significant effect on test performance when choosing the right
learning rate. The use of a small learning rate will require a small batch size to produce maximum accuracy.
The greater the epoch value, the higher the accuracy value generated, the greater the accuracy value will result
in a smaller loss value and vice versa. Based on the experiments that have been carried out, the researcher
concludes that the improvement in model performance is not only influenced by hyperparameter optimization
but also by image preprocessing. Hyperparameter optimization is able to improve model performance, but the
accuracy value obtained is less than optimal. Therefore, researchers perform image preprocessing to improve
the results of model performance. The data used in this study has been processed in such a way as described in
the image preprocessing design.

ACKNOWLEDGEMENTS
This work was supported and funding by the research institutions and community service of Institut
Teknologi Del.

REFERENCES
[1] A. D. Ware, “The complete blood count and white blood cell differential,” Contemporary Practice in Clinical Chemistry,
pp. 429–444, 2020, doi: 10.1016/b978-0-12-815499-1.00025-9.
[2] K. Throngnumchai, P. Lomvisai, C. Tantasirin, and P. Phasukkit, “Classification of white blood cell using deep
convolutional neural network,” BMEiCON 2019-12th Biomedical Engineering International Conference, 2019,
doi: 10.1109/BMEiCON47515.2019.8990301.
[3] M. Yildirim and A. Çinar, “Classification of white blood cells by deep learning methods for diagnosing disease,” Revue
d’Intelligence Artificielle, vol. 33, no. 5, pp. 335–340, 2019, doi: 10.18280/ria.330502.
[4] K. Kousalya, S. Santhiya, K. Dinesh, R. S. Mohana, and B. Krishnakumar, “Blood cells classification using convolutional neural
network architecture,” International Journal of Advanced Science and Technology, vol. 29, no. 3 Special Issue, pp. 261–267, 2020.
[5] K. Dong, C. Zhou, Y. Ruan, and Y. Li, “MobileNetV2 model for image classification,” Proceedings-2020 2nd
International Conference on Information Technology and Computer Application, ITCA 2020, pp. 476–480, 2020,
doi: 10.1109/ITCA52113.2020.00106.
[6] R. R. N. M. I. Tobias et al., “CNN-based deep learning model for chest X-ray health classification using TensorFlow,” Proceedings-
2020 RIVF International Conference on Computing and Communication Technologies, RIVF 2020, 2020,
doi: 10.1109/RIVF48685.2020.9140733.
[7] N. Yuningsih and M. Mustikasari, “Anemia classification based on abnormal red blood cell morphology using convolutional neural
network,” IOSR J. Comput. Eng., vol. 22, no. 1, pp. 22–29, 2020, doi: 10.9790/0661-2201042229.

Image preprocessing and hyperparameter optimization on pretrained model … (Parmonangan R. Togatorop)

1222  ISSN: 2252-8938

[8] P. M. Radiuk, “Impact of training set batch size on the performance of convolutional neural networks for diverse datasets,”
Information Technology and Management Science, vol. 20, no. 1, 2018, doi: 10.1515/itms-2017-0003.
[9] N. Xu, “Research on the influence of convolutional neural network parameters on fruit classification,” 2019 IEEE 3rd International
Conference on Electronic Information Technology and Computer Engineering, EITCE 2019, pp. 606–610, 2019,
doi: 10.1109/EITCE47263.2019.9094902.
[10] P. Probst, A. L. Boulesteix, and B. Bischl, “Tunability: Importance of hyperparameters of machine learning algorithms,” Journal
of Machine Learning Research, vol. 20, 2019.
[11] M. Shakunthala and K. Helen Prabha, “Preprocessing analysis of brain iimages with atherosclerosis,” Proceedings of 2019 3rd
IEEE International Conference on Electrical, Computer and Communication Technologies, ICECCT 2019, 2019,
doi: 10.1109/ICECCT.2019.8869009.
[12] F. Novoselnik, R. Grbic, I. Galic, and F. Doric, “Automatic white blood cell detection and identification using convolutional neural
network,” Proceedings of International Conference on Smart Systems and Technologies 2018, SST 2018, pp. 163–167, 2018,
doi: 10.1109/SST.2018.8564625.
[13] K. Bousbai and M. Merah, “A comparative study of hand gestures recognition based on MobileNetV2 and ConvNet models,”
Proceedings - 2019 6th International Conference on Image and Signal Processing and their Applications, ISPA 2019, 2019,
doi: 10.1109/ISPA48434.2019.8966918.
[14] M. Sandler, A. Howard, M. Zhu, A. Zhmoginov, and L. C. Chen, “MobileNetV2: Inverted residuals and linear bottlenecks,”
Proceedings of the IEEE Computer Society Conference on Computer Vision and Pattern Recognition, pp. 4510–4520, 2018,
doi: 10.1109/CVPR.2018.00474.
[15] A. G. Howard et al., “MobileNets: Efficient convolutional neural networks for mobile vision applications,” 2017,
[Online]. Available: http://arxiv.org/abs/1704.04861.
[16] R. Patel and A. Chaware, “Transfer learning with fine-tuned MobileNetV2 for diabetic retinopathy,” 2020 International Conference
for Emerging Technology, INCET 2020, 2020, doi: 10.1109/INCET49848.2020.9154014.
[17] C. Buiu, V. R. Dănăilă, and C. N. Răduţă, “MobileNetV2 ensemble for cervical precancerous lesions classification,” Processes,
vol. 8, no. 5, 2020, doi: 10.3390/PR8050595.
[18] N. H. Barnouti, “Improve face recognition rate using different image pre-processing techniques,” American Journal of Engineering
Research (AJER), vol. 5, no. 4, pp. 46–53, 2016, [Online]. Available: www.ajer.org.
[19] S. Genc, K. N. Akpinar, and S. Karagol, “Automated abnormality classification of chest radiographs using MobileNetV2,” HORA
2020 - 2nd International Congress on Human-Computer Interaction, Optimization and Robotic Applications, Proceedings, 2020,
doi: 10.1109/HORA49412.2020.9152607.
[20] K. Thenmozhi and U. S. Reddy, “Image processing techniques for insect shape detection in field crops,” Proceedings of
the International Conference on Inventive Computing and Informatics, ICICI 2017, pp. 699–704, 2018,
doi: 10.1109/ICICI.2017.8365226.
[21] J. Shijie, W. Ping, J. Peiyi, and H. Siping, “Research on data augmentation for image classification based on convolution neural
networks,” Proceedings-2017 Chinese Automation Congress, CAC 2017, vol. 2017-January, pp. 4165–4170, 2017,
doi: 10.1109/CAC.2017.8243510.
[22] C. C. Aggarwal, Neural Networks and Deep Learning. Cham: Springer International Publishing, 2018.
[23] Z. Gao and X. Wang, “Deep learning,” EEG Signal Processing and Feature Extraction, pp. 325–333, 2019, doi: 10.1007/978-981-
13-9113-2_16.
[24] G. Montavon, G. B. Orr, and K.-R. Mueller, “Neural networks: Tricks of the trade - Second Edition,” Springer, pp. 639–655, 2012.
[25] M. J. Macawile, V. V. Quiñones, A. Ballado, J. Dela Cruz, and M. V. Caya, “White blood cell classification and counting
using convolutional neural network,” 2018 3rd International Conference on Control and Robotics Engineering, ICCRE 2018,
pp. 259–263, 2018, doi: 10.1109/ICCRE.2018.8376476.

BIOGRAPHIES OF AUTHORS

Parmonangan R. Togatorop Master's in information technology, Bachelor in

information system, and Faculty Members & Researcher in Institut Teknologi Del. Research
area in artificial intelligence, enterprise system, and data warehouse. She can be contacted at
email: [email protected]

Yohanssen Pratama Current Faculty Members & Researcher in Institut

Teknologi Del. 4+ years experience specializing in back-end/infrastructure, analytical tools
development and computer programming. Teach academic and vocational subjects to
undergraduates and also pursue my own research to contribute to the wider research activities
of my department. He also a doctoral candidate in Nara Institute of Science and Technology
and can be contacted at email: [email protected] or
[email protected]

Int J Artif Intell, Vol. 12, No. 3, September 2023: 1210-1223

Int J Artif Intell ISSN: 2252-8938  1223

Astri Monica Sianturi Currently working as a Software Quality Assurance in

PT. Phincon which is an IT consultant in Jakarta focusing on middleware and CRM. Have
a professional experience of 2+ years in developing software and using development tools.
She can be contacted at email: [email protected]

Mega Sari Pasaribu Currently working as a Software Quality Assurance in PT.

Phincon which is an IT consultant in Jakarta focusing on middleware and CRM. Have a
professional experience of 2+ years in developing software and using development tools. She
can be contacted at email: [email protected]

Pebri Sangmajadi Sinaga Currently working as a Software Engineer in PT.

Tokopedia specific in Engineering Productivity which is one of the biggest e-commerce in
Indonesia. Have a professional experience of 2+ years in developing software and using
development tools. He can be contacted at email: [email protected]

Image preprocessing and hyperparameter optimization on pretrained model … (Parmonangan R. Togatorop)