Deep Learning-Based Convolutional Neural Network for Flower

Classification

Yangchuan Liu

High School Affiliated to South China Normal University, Guangzhou, China

Keywords: Machine Learning, Convolutional Neural Network, Flower Classification.

Abstract: To tackle the flower classification problem, this study utilizes the Oxford 102 Flowers dataset and develops

a machine learning model using a Convolutional Neural Network (CNN). Initially, the preprocessing phase

involved the use of grayscale images. However, recognizing the critical role that color plays as a distinctive

feature of flowers, the study shifted to using RGB images. To further enhance the model's performance, data

augmentation techniques were introduced. These included random adjustments to brightness, saturation,

contrast, and hue, which helped diversify the training set and improve the model's generalization ability. To

mitigate overfitting, several strategies were employed, such as tuning the number of neurons and incorporating

Dropout Layers. These approaches helped the model achieve a validation accuracy of approximately 0.7,

which is sufficiently accurate for basic flower classification tasks in everyday applications. This outcome

demonstrates the effectiveness of the chosen methods and highlights the potential of CNNs in flower image

classification.

1 INTRODUCTION

The flower is the reproductive organ of a seed plant.

Composed of corolla, calyx, receptor, and stamen, it

has a variety of shapes, colors, and fragrances. There

are countless types of flowers worldwide, which have

multiple functions in different domains. For example,

flowers can be used as decorations in ceremonial

events, given to others as gifts on special days, and

even processed into food or tea. Because of these

uses, there is a considerable flower demand that leads

to a large market for the flower flower-growing

industry, and technology must be the key to

improving the productivity of flowers. However,

flower classification is one of the challenging tasks in

flower growing. Traditional flower identification

methods rely on specialized knowledge and manual

classification, which can be time-consuming, labor-

intensive, and subject to personal bias. With the rapid

advancement of computer vision and machine

learning technologies, data-driven automated

recognition systems have emerged as a solution to

improve productivity and reduce labor costs.

In recent years, Artificial Intelligence (AI)

technology has gradually matured, and many fields

https://orcid.org/0009-0000-0082-2061

have been widely used. In the manufacturing

industry, image recognition technology and computer

vision offer developing opportunities for driverless

cars and automatic face recognition systems. In the

economy, machine learning provides a more reliable

way to predict the future market of certain goods by

training with past data. In medicine, neural

networking can precisely produce predictions about

patients’ latent diseases by collecting current body

data. More generally, the current Large Language

Model (LLM), such as Chat Generative Pre-Trained

Transformer (ChatGPT), can even provide help and

opinions to people in all fields. There is undoubtedly

an astonishing advancement in the AI domain. For

instance, back to focus on the planting industry, in

2016, Liu et al. proposed a novel framework based on

a Convolutional Neural Network (CNN) that

achieves 84.02% flower classification accuracy (Liu,

2016). Based on the study of Knauer et al.in 2019,

they found that CNN has a better performance in the

tree species classification task than the Random

Forest (RF) classifiers (Knauer, 2019). By analyzing

the previous relative research and confirming the

effectiveness of AI, this research aims to consider the

machine learning technology based on CNN to be

Liu, Y.

Deep Learning-Based Convolutional Neural Network for Flower Classiﬁcation.

DOI: 10.5220/0013329800004558

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 1st International Conference on Modern Logistics and Supply Chain Management (MLSCM 2024), pages 315-319

ISBN: 978-989-758-738-2

315

feasible and suitable to solve the problem of flower

classification.

To achieve the final goal, this paper first decided

and downloaded a suitable dataset Oxford 102

Flowers that contains 8, 200 samples of 102 different

types of flowers. Then this study standardized the

image by resizing all of them to 200-pixel length

squares and converting them into grayscale. The

model’s neural network consists of 4 convolutional

layers and 4 dense layers, and all of them use

Rectified Linear Unit (ReLu) as the activation

function. The input data is the pixel of an image, and

the output is a vector that reflects the possibility of

each of the 102 types. In the training process, the

Adaptive Moment Estimation (Adam) was used as an

optimizer, and the Cross Entropy Error was used as a

loss function. After training of 100 epochs, the model

accuracy finally reaches about 70%.

2 METHOD

2.1 Dataset Preparation

In the dataset preparation part, Oxford 102 Flower

(102 Category Flower Dataset) is chosen as the

training dataset in this study (Nilsback, 2008). This

dataset consists of 8,200 RGB images divided into

102 flower categories that are commonly occurring in

the United Kingdom. Each class contains between 40

and 258 images, and the images have large scale,

pose, and light variations. Including these diverse

and complicated features, this dataset is suitable for

the model that aims to solve the complex flower

classification task. When preprocessing the images,

the grayscale images were first considered an

appropriate method. As the RGB images are

translated into the grayscale, image noise is reduced,

while texture and structural features are enhanced,

thus increasing the efficiency of model processing.

Figure 1 clearly shows the difference between the

RGB images and grayscale images.

Figure 1: The RGB version(right) and the grayscale

version(left) of a flower image (Photo/Picture credit:

Original).

However, the study soon found that this

preprocessing method leads to a low accuracy

because the color is a necessary indicator of the

flower categories. As a result, the gray-scale images

are not feasible in the flower classification task;

instead, considering the importance of the color,

random adjustments of image brightness, saturation,

contrast, and hue were used to enhance the model's

adaptability to color changes. In addition, random

up-and-down or side-to-side flips also improve the

generalization of the model, and all of the images are

resized into 200 pixels length squares, as normalizes

the input shape of the model.

2.2 Convolutional Neural

Network-Based Prediction

In the step of building up the neural network model,

the Convolutional Neural Network (CNN) is first to

be considered. CNN is predominantly used to extract

the features from the grid-like matrix dataset, such as

visual datasets like images or videos (Li, 2021; Gu,

2018; O'Shea, 2015). The components in the CNN

include the Input Layer, Convolutional Layer,

Activation Layer, Pooling Layer, Flattening, Fully

Connected Layer, and Output Layer. In these layers,

the Convolutional Layer and Pooling Layer are the

essences that make CNN different from other models.

Imagine that an image is a cuboid; specifically, the

width and length of the cuboid are those of the image

while the channels(height) can represent the RGB

value of each pixel. By taking a small patch of this

cuboid and running a small neural network, which is

called a filter, the Convolutional Layer can extract the

feature of an image when the filter slides in the cuboid

and convert them into another image with different

width, height, and channels, which is referred as

feature maps. As important as the Convolutional

Layer, the Pooling Layer is periodically inserted

between the Convolutional Layer to reduce the size of

volume which makes the computation fast reduces

memory produced by multiple feature maps and

prevents overfitting. Because the advantages of CNN

include high accuracy of image analysis, robustness

to image deformation and rotation, and the need for a

large amount of label data, this network is suitable for

the flower classification task that meets these

conditions.

Considering the complex data contained in each

image, the study uses 4 Convolutional Layers, 4

Pooling Layers, and 3 Fully Connected Layers to

construct the model. Nevertheless, in this version, the

model still does not perform well after 100 epochs.

By studying the flaws of the traditional CNN, the

MLSCM 2024 - International Conference on Modern Logistics and Supply Chain Management

316

Figure 2: Model accuracy when use the gray-scale images in the preprocessing (Photo/Picture credit: Original).

Figure 3: Model accuracy when using the RGB image and data augmentations in the preprocessing (Photo/Picture credit:

Original).

Figure 4: Model accuracy when adding some Dropout Layers in the model (Photo/Picture credit: Original).

study notices that overfitting is the main cause of

low performance. Overfitting always happens when

fully connected layers have a large number of neurons

that extract very similar features from the input data;

then, these neurons add more significance to those

features for the model, which only work in specific

datasets (Srivastava, 2013; Salakhutdinov, 2014). To

solve this problem, Dropout Layers -- randomly shut

down some fraction of a layer’ s neurons by zeroing

out the neuron values -- are added between each Fully

Connected Layer to reduce the model ’ s over-

relying on specific neurons.

2.3 Implementation Details

The hyper-parameters of the model are also

significant, such as the optimizer, loss function,

neuron amount, learning rate, and so on. Because the

task is a classification problem, Cross Entropy

Error is used as the loss function (Golik, 2013).

Based on the complexity of the flower classification

task, Adaptive Moment Estimation (Adam) (Tato,

2018; Zhang, 2018), which automatically adjusts the

learning rate and occupies lower memory, is adopted

as the optimizer. Moreover, to capture abundant

Deep Learning-Based Convolutional Neural Network for Flower Classiﬁcation

317

features of flowers, the model has a relatively large

number of neurons in Fully Connected Layers and

filters in Convolutional Layers.

3 RESULTS AND DISCUSSION

In Figure 2, the graph shows the validation accuracy

and training accuracy of the model that uses the gray-

scale images in the preprocessing. It is clear that the

training accuracy increases rapidly during the training

process and even reaches over 0.9 after 20 epochs.

However, the validation accuracy remains at a

shallow level, basically no more than 0.1.

In the Figure 3, RGB images are used instead of

gray-scale images; moreover, data augmentations are

also added in the preprocessing. The graph shows

that the validation accuracy begins to grow with the

training accuracy and fits well in the early stage.

However, after 40 epochs, the two lines begin to

separate: the difference between the training accuracy

and validation is greater and greater. At the same

time, the final validation accuracy, which stops

growing and reaches only about 0.4, is not enough for

the expectation of the study.

In the Figure 4, Dropout Layers are added and the

neuron amount in some layers is adjusted in the

model. In the graph, the two lines fit well in general,

and the final validation accuracy makes a huge

improvement, which reaches nearly 0.7.

Figure 5: Model’ s prediction of two image with data

augmentation (Photo/Picture credit: Original).

As shown in Figure 5, both of the images

experience data augmentation that randomly changes

brightness, saturation, contrast, and hue significantly.

The prediction of the first one can still be correct

while that of the second one is incorrect.

From the information of the first version model,

because it only takes the model 20 epochs to reach the

training accuracy over 0.9, it is easy to find the high

efficiency of gray-scale preprocessing in image-

related tasks. Nevertheless, because flower color as a

feature cannot be ignored, the gray-scale image is not

feasible. Some inappropriate features, which only

belong to the training dataset and do not exist in the

validation dataset, are extracted and trained by the

model. As a result, although the training accuracy is

impressive, the validation accuracy is extremely low,

which does not meet the requirement.

In the second version, the RGB images and data

augmentations help the model begin to capture and

analyze the features of flowers’color. The fitting of

training accuracy and validation accuracy proves the

opinion that color is non-negligible. However, in this

step, the increased complexity due to converting 1-

channel grayscale images to 3-channel RGB images

is not considered. Therefore, the simple model could

not adapt to the abundant features carried by the RGB

images and did not perform well, only reaching a

validation accuracy of about 0.4 before it stopped

growing accuracy.

In the final version, the neuron amount is adjusted

to improve the model complexity, while the Dropout

Layers are used to decrease redundancy features

recorded in neurons to avoid overfitting. Because of

the improvement of the complexity, the model can

afford and extract the features well and finally

reaches its peak validation accuracy of about 0.7

without an obvious overfitting phenomenon. It is

undoubtedly precise enough for some basic flower

classification tasks in daily life because most flower

images in real life have more distinguishing features

and do not resemble the images with data

augmentations in Figure 4, which could be even

harder for human beings to discern.

However, there are still some limitations of the

proposed model including lacking training data: the

8,200 images used to classify 102 kinds of flowers are

not enough. To achieve higher accuracy, by

considering CNN’s reliance on the dataset, more

images should be collected as the training dataset of

this model.

4 CONCLUSIONS

In this study, a machine learning model is built and

trained to solve the flower classification task. Oxford

102 Flower which consists of 8,200 RGB images

divided into 102 flower categories is used as the

dataset. For its high performance in image-related

tasks, CNN is considered the most suitable network

of the model instead of Random Forest (RF);

furthermore, to improve its accuracy and robustness,

MLSCM 2024 - International Conference on Modern Logistics and Supply Chain Management

318

data augmentations, large number of neurons, and

Dropout Layers are used. As a result, the accuracy of

the model reaches about 0.7, which is impressive and

enough for most practical flower classification tasks

when it has to classify some images with terrible

distortion. Some improvements can be made in the

future if use larger and more diverse datasets to

improve the accuracy to a higher level.

REFERENCES

Gu, J., Wang, Z., Kuen, J., Ma, L., Shahroudy, A., Shuai,

B., ... & Chen, T. 2018. Recent advances in

convolutional neural networks. Pattern recognition, 77,

354-377.

Golik, P., Doetsch, P., & Ney, H. 2013, August. Cross-

entropy vs. squared error training: a theoretical and

experimental comparison. In Interspeech (Vol. 13, pp.

1756-1760).

Knauer, U., von Rekowski, C. S., Stecklina, M., Krokotsch,

T., Pham Minh, T., Hauffe, V., ... & Seiffert, U. 2019.

Tree species classification based on hybrid ensembles

of a convolutional neural network (CNN) and random

forest classifiers. Remote Sensing, 11(23), 2788.

Li, Z., Liu, F., Yang, W., Peng, S., & Zhou, J. 2021. A

survey of convolutional neural networks: analysis,

applications, and prospects. IEEE transactions on

neural networks and learning systems, 33(12), 6999-

7019.

Liu, Y., Tang, F., Zhou, D., Meng, Y., & Dong, W. 2016,

November. Flower classification via convolutional

neural network. In 2016 IEEE International Conference

on Functional-Structural Plant Growth Modeling,

Simulation, Visualization and Applications (FSPMA)

(pp. 110-116). IEEE.

Nilsback, M. E., & Zisserman, A. 2008, December.

Automated flower classification over a large number of

classes. In 2008 Sixth Indian conference on computer

vision, graphics & image processing (pp. 722-729).

IEEE.

O'Shea, K. 2015. An introduction to convolutional neural

networks. arXiv preprint arXiv:1511.08458.

Srivastava, N. 2013. Improving neural networks with

dropout. University of Toronto, 182(566), 7.

Srivastava, N., Hinton, G., Krizhevsky, A., Sutskever, I., &

Salakhutdinov, R. 2014. Dropout: a simple way to

prevent neural networks from overfitting. The journal

of machine learning research, 15(1), 1929-1958.

Tato, A., & Nkambou, R. 2018. Improving adam optimizer.

Openreview

Zhang, Z. 2018, June. Improved adam optimizer for deep

neural networks. In 2018 IEEE/ACM 26th international

symposium on quality of service (IWQoS) (pp. 1-2).

IEEE.

Deep Learning-Based Convolutional Neural Network for Flower Classiﬁcation

319