Fruit Image Classification Based on SVM, Decision Tree and KNN

Xiang Han

Faculty of Information Science and Engineering, Ocean University of China, Qingdao, China

Keywords: Fruit Image Classification, SVM, Decision Tree, KNN.

Abstract: Image classification is becoming more and more popular in today’s daily life. Image classification is widely

used in many fields. For example, the market demand for face recognition technology has increased

significantly in recent years. The foundation of these new technologies is still image classification. In order

to explore the efficiency of different image classification algorithms and help guide the use of different image

classification algorithms in the market, this paper uses a variety of algorithms to classify images in the fruit360

dataset. Fruit360 dataset is a dataset with 90483 images of 131 kinds of fruits and vegetables. Images in this

dataset have a size of 100×100 pixels. As a result, the Support Vector Machine algorithm is 89% accurate,

the decision tree algorithm is 94% accurate, and the K-nearest Neighbors algorithm is nearly 100% accurate.

Apart from the accuracy of these algorithms, this paper also analyzes the difference in classification accuracy

among different classes. For the Support Vector Machine algorithm, the classification accuracy of class 1 and

class 2 is low, which is caused by the algorithm itself. For the decision tree algorithm, the accuracy of each

classification group is similar. For the K-nearest Neighbors algorithm, the overall accuracy is very high. In

addition, this paper also compares the characteristics of these three algorithms, analyzes the performance

difference between the Support Vector Machine algorithm and the decision tree algorithm, and discusses the

relationship between the Support Vector Machine algorithm efficiency and the number of classes.

1 INTRODUCTION

Nowadays, image classification is widely used in

people’s daily life. Taking the campus as an example,

schools need to use a face recognition system at the

entrance of the library to help determine whether

there is access permission. But the author sometimes

sees such cases: some people outside the school can

also successfully pass the face recognition system.

This means that the current image classification still

has the problem of low accuracy. In this paper, SVM,

decision tree and KNN, three different algorithms are

used, and based on fruit360 dataset, various

classification algorithms are compared and analyzed,

so as to give a reasonable method selection for image

classification problems.

There have been many scientific studies based on

these algorithms. Boumedine Ahmed Yassine et al.

used KNN for 3D face recognition. They used KNN

for feature extraction (Yassine et al 2023). In other

research, they did glass component classification

based on decision tree (Guo et al 2023). However,

most of the existing research lack the comparison

between algorithms.

In addition, there are many research on image

classification using neural networks (Zhang et al

2014). However, a neural network is an end-to-end

model, which remains a “black box” for users, and

people cannot intuitively see the operation process of

its internal algorithms (Wang et al 2020)0. Also, the

training time of neural network algorithm is long, and

the interpretation is not strong enough. So this paper

focuses on the SVM, decision tree and KNN

algorithm.

Support Vector Machine is a supervised machine

learning algorithm, and it is widely used for

regression and classification tasks (Chandra and Bedi

2018 & Nie et al 2023). In addition to image

classification, SVM can be applied to text

classification, such as sentiment analysis, topic

classification, etc. SVM can extract text features and

make decisions for classification. SVM can also be

used for financial forecasting. In the financial field,

SVM can be used to predict stock prices and calculate

the risk of investments.

The decision tree algorithm is used in the financial

field and medical field. In the financial field, banks

can use a decision tree model to predict a customer’s

Han, X.

Fruit Image Classiﬁcation Based on SVM, Decision Tree and KNN.

DOI: 10.5220/0012815800003885

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

In Proceedings of the 1st International Conference on Data Analysis and Machine Learning (DAML 2023), pages 367-373

ISBN: 978-989-758-705-4

367

credit rating by taking factors such as a customer’s

personal information and historical loan history. In

the medical field, the decision tree algorithm can be

used for disease diagnosis, drug therapy, etc. For

doctors, they can use decision tree models to predict

the situation of a patient by taking plenty of factors

such as symptoms and other indexes.

A machine learning approach called K-nearest

neighbors may be applied to tasks involving

regression and classification (Soucy and Mineau

2002). For speech recognition, it can classify audio

signals into different speech or voice types by

analyzing and comparing those given data. As long as

the scene meets the standard classification or

regression problem, it can try to use the KNN

algorithm to solve it. However, it should be noted that

the main disadvantage of the KNN algorithm is high

computational complexity, especially when the data

set is large, and the calculation distance requires more

time and computing resources.

In this paper, the author uses various methods,

including SVM, decision tree and KNN, to finish this

image classification task. By using different

algorithms for image classification, the differences of

different algorithms on this task would appear. Using

a variety of evaluation indexes to estimate the

efficiency of different algorithms.

2 METHOD

2.1 Dataset Analysis

In this section, the author will have a brief introduction

to the dataset. The name of this dataset is “fruit-360”,

which means that it has a large quantity of fruit

images. So here, the author first calculates the amount

of images in the folder. There are 90483 images in the

dataset. The training set contains 67692 images and

the test set contains 22688 images. There are 131

classes of fruits. Fig. 1. shows some pictures in the

dataset in Jupyter Notebook.

To simplify the problem, only 10 classes of fruits

were used for image classification. Doing this can

reduce the expense of time, which can highly develop

the efficiency of evaluating different algorithms.

2.2 Data Preprocessing

By fetching images from the given folder, the author

imports pictures into Python for further classification.

Using Python to get those folders’ names and list them

in an array. Then, with a random rate of 100%, the

original training data are divided into 2 sets. The

training set accounts for 80% and the test set accounts

for 20%.

Figure 1: Some Pictures in the Dataset (Picture credit:

Original).

After that, by resizing images, they are adjusted to

the same scale. Then calculate the image histogram for

further preparation.

2.3 Support Vector Machine

Support Vector Machine is a data-based classification

algorithm, which focuses on establishing a hyperplane

to divide data.

By finding the best hyperplane in space, SVM

could reach a result that the distance between positive

samples and negative samples is the farthest. In binary

classification, SVM could be a line in a 2-

dimensionality figure. It can also be used in a n-

dimensionality question, where n is larger than 2. Fig.

2 shows the utilization in solving binary classification

problems (Guo et al 2023) and Fig. 3. shows the

utilization in solving 3-dimensionality problems.

Figure 2: SVM Principle (2-dimensionality).

DAML 2023 - International Conference on Data Analysis and Machine Learning

368

Figure 3: SVM Principle (3-dimensionality).

2.4 Decision Tree

The decision tree algorithm is a method of

approximating discrete function values. It is an

algorithm with a tree structure, and each point means

an attribute for judgements. Leaves refer to a

classification result. By using decisions, the decision

tree algorithm can analyze data using readable rules.

Figure 4: Train of Thought of Decision Tree (Picture credit:

Original).

The decision tree algorithm is a common

supervised learning algorithm. By determining

whether the input data has a label, people can know

whether it is supervised learning. If the input data is

labeled, it is supervised learning. In this task, the data

set is given labels for different pictures, which can be

seen as supervised learning. With supervised learning,

the model can use previous experience for evaluating

the output result, and the accuracy of prediction can be

high.

The train of thought of decision tree is shown in

Fig. 4.

2.5 K-Nearest Neighbors

For classification problems, K-Nearest Neighbors is

an efficient technique that may be applied broadly.

The K-nearest data points may be used to forecast a

specific query point, and the labels of these neighbors

can then be used to predict the query point. In

classification tasks, the predicted label for query

points is determined by the weight of those K-nearest

neighbors. The KNN algorithm can also be used for

text classification. By extracting the features of words,

the text is classified according to different dimensions

(Li 2021).

2.6 Evaluation

By calculating the value of Accuracy, Precision,

Recall and F1-score, the rationality of the algorithm

would be evaluated. Also, by using the test dataset

and calculating the probability of correct

classifications, the model can be proved to be

practical or not. During programming, by

automatically generating the result of Precision,

Recall and F1-score, the accuracy of a certain model

can be found, which is helpful for evaluation.

Confusion matrix is an important tool for

supervised learning. It is a form, which contains rows

representing the practical classes and columns

representing the predicted classes. Each cell

expresses the frequency in a certain state. By

analyzing the confusion matrix, the accuracy of the

model will be further indicated.

Accuracy reflects the proportion of those true

predictions for positive and negative examples. When

a model has a high level of accuracy, it means that the

amount of correct prediction is large. Using (1) can

calculate the value of accuracy.

Precision can reflect the proportion of accurate

predictions based on all predictions. It shows the

proportion of actual predictions in all positive

samples. Equation (2) shows the process of

calculating the value of Precision.

The process of calculating Recall can be explained

by (3).

F1-score is a comprehensive evaluation index for

evaluation, which is based on Precision and Recall. If

a model has a high F1-score, it would be balanced,

which means that one of the values of Precision and

Recall would not be too high. F1-score can be

calculated by (4).

𝐴

𝑐𝑐𝑢𝑟𝑎𝑐𝑦 =

𝑇𝑁 + 𝑇𝑃

𝑇𝑁 + 𝑇𝑃 + 𝐹𝑁 + 𝐹𝑃





𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛 =

𝑇𝑃

𝑇𝑃 + 𝐹𝑃





𝑅𝑒𝑐𝑎𝑙𝑙 =

𝑇𝑃

𝑇𝑃 + 𝐹𝑁





Fruit Image Classiﬁcation Based on SVM, Decision Tree and KNN

369

3 RESULT

3.1

Runtime Environment

The environment for these algorithm is Jupyter

Notebook 6.6.4. All the code is run on a Dell Inspiron

7400 laptop, with 11

Gen Intel® Core™ i7-1165G7

@ 2.80GHz CPU and Intel® Iris® Xe Graphics GPU.

3.2 SVM Result

Using SVM for image classification, and results are

shown in Table 1.

Table 1: SVM Result.

No. Precision Recall F1-score Support

0 0.52 0.58 0.55 106

1 0.69 0.71 0.70 87

2 1.00 1.00 1.00 99

3 1.00 1.00 1.00 91

4 1.00 1.00 1.00 87

5 1.00 1.00 1.00 92

6 1.00 0.97 0.98 92

7 0.99 0.98 0.98 100

8 0.94 0.84 0.89 113

9 0.89 0.88 0.88 83

accuracy - - 0.89 950

macro avg 0.90 0.90 0.90 950

weighted avg 0.90 0.89 0.90 950

The confusion matrix is shown in Fig. 5.

Figure 5: Confusion Matrix for SVM (Picture credit:

Original).

Code execution time: 386.26s.

The SVM algorithm requires a long time to finish

this image classification challenge. There are some

reasons for this:

• When the scale of the dataset is large, the distance

between each sample point and all other sample

points needs to be calculated, which causes the

computational complexity to increase

dramatically as the number of samples increases.

• When dealing with non-linear problems, SVM

requires kernel functions to map primitive

features into higher-dimensional Spaces. This

process can lead to computational complexity,

making the training process slow.

• Binary classification problems were the initial

purpose of SVM's creation. Various binary

classifications are necessary when handling

various classification problems. This process can

lead to increased computational complexity,

making the training process slower. This aspect

will be discussed in the discussion part of this

paper.

There are some characteristics in Fig. 5. In Fig. 5.,

class 0 and class 1’s results are not as good as

expected. Some pictures in class 1 are classified as

class 2, and some pictures in class 2 are classified as

class 1. It is inferred that there is little difference

between class 1 and class 2 in image data features,

which leads to an inaccurate result. Also, due to the

fact that images from class 0 and class 1 are quite

similar to images from class 8 and class 9, which

leads to misclassification.

3.3 Decision Tree Result

Using the decision tree algorithm for image

classification, the results are shown in Table 2.

Table 2: Decision Tree Result.

No. Precision Recall F1-score Support

0 0.94 0.84 0.89 106

1 0.96 0.98 0.97 87

2 0.95 0.98 0.97 99

3 0.98 0.95 0.96 91

4 0.94 0.95 0.95 87

5 0.97 0.95 0.96 92

6 0.95 0.95 0.95 92

7 0.91 0.96 0.94 100

8 0.92 0.93 0.93 113

9 0.93 0.98 0.95 83

accuracy - - 0.94 950

macro avg 0.94 0.95 0.94 950

weighted avg 0.94 0.94 0.94 950

𝐹

𝑃𝑟𝑒𝑐𝑖𝑠𝑖𝑜𝑛

𝑅𝑒𝑐𝑎𝑙𝑙





DAML 2023 - International Conference on Data Analysis and Machine Learning

370

The confusion matrix is shown in Fig. 6.

Figure 6: Confusion Matrix for Decision Tree (Picture

credit: Original).

Code execution time: 25.34s.

In Fig. 6., the accuracy of different classes is very

close. There is no longer a phenomenon that the

classification accuracy of some groups is very low as

in the results obtained by the SVM algorithm.

Consider that this effect is caused by image noise.

Superfluous information in photographs is

referred to as image noise. Image noise can be defined

as any variety of elements present in image data that

may impede individuals from making accurate

decisions. Image noises are random and

unpredictable. Usually, while acquiring a picture,

image noise data are produced. The working

environment is often the cause of image noise. Circuit

construction and electronic components may also be

to blame. Image noise can occasionally be produced

as a result of pollution that occurs during media

transmission and recording equipment.

In Fig. 6., each group has a margin of error, which

means they have something in common. Due to the

limited number of image pixels in the dataset, the

prediction accuracy cannot be further improved.

3.4 KNN Result

Using KNN for image classification, the results are

shown in Table 3.

Table 3: KNN Result.

No. precision recall f1-score support

0.99

1.00 1.00 106

1 1.00 0.99 0.99 87

2 1.00 1.00 1.00 99

No. precision recall f1-score support

3 1.00 1.00 1.00 91

4 1.00 1.00 1.00 87

1.00

1.00 1.00 92

0.99

1.00 0.99 92

7 1.00 1.00 1.00 100

8 1.00 1.00 1.00 113

9 1.00 0.99 0.99 83

accuracy - - 1.00 950

macro avg 1.00 1.00 1.00 950

weighted avg 1.00 1.00 1.00 950

The confusion matrix is shown in Fig. 7.

Figure 7: Confusion Matrix for KNN (Picture credit:

Original).

Code execution time: 14.20s.

In Fig. 7., the accuracy of KNN is quite high. In

the experiment, merely a few data have wrong

prediction results. KNN can handle classification

problems, and it can naturally handle multiple

classification problems compared with SVM. The

accuracy of the KNN algorithm is improved by

finding the appropriate K value through loop

traversal. Therefore, the classification results of the

KNN algorithm can be interpreted more accurately

from the perspective of its characteristics.

3.5 Classification Result Evaluation

There is a significant variation in algorithm efficiency

in Fig. 8. SVM is more time-consuming than KNN

and decision trees. The SVM algorithm's

performance generally declines as the number of

classes increases. It was initially created to solve

binary classification problems.

Fruit Image Classiﬁcation Based on SVM, Decision Tree and KNN

371

Figure 8: Classification Runtime (Picture credit: Original).

Compared to SVM and the decision tree

algorithm, SVM needs to find the optimal hyperplane

to distinguish data during training, which requires

repeated calculations until a hyperplane satisfying the

conditions is found. It can take a long time to

complete this procedure, particularly when working

with big datasets. In contrast, in the training stage, the

decision tree algorithm simply uses a rectangle to

divide the feature space, so the decision tree

algorithm has a relatively small calculation amount.

In contrast to SVM and KNN, KNN requires less

computation and operates more quickly because it

just has to store training data during the training stage

and compute the distance between the sample to be

predicted and the training data during the prediction

stage.

3.6 Results for SVM Algorithm

Efficiency

In order to find out the law between code execution

time and the amount of classes, the author did further

research on efficiency. By adjusting the amount of

classes from 2 to 10, the result is shown in Table 4.

Table 4: SVM Runtime and Accuracy.

Num of classes Runtime Accuracy

2 7.73 0.97

3 17.43 0.92

4 53.00 0.91

5 65.34 0.91

6 117.26 0.90

7 147.65 0.90

8 282.63 0.91

9 323.07 0.91

10 356.27 0.89

4 DISCUSSION

4.1 Accuracy Differences Between

SVM and Decision Tree

This may caused by the characteristics of SVM and

decision tree. Image data usually has a large number

of features, and the relationships between features can

be very complex. SVM may have difficulty in dealing

with high-dimensional features, especially when

dealing with non-linear problems. In contrast, decision

trees may be more flexible when dealing with

nonlinear problems.

4.2 SVM Algorithm Operation

Efficiency

Figure 9: SVM Algorithm Operation Efficiency (Picture

credit: Original).

In Table 4 and Fig. 9., the tendency of runtime and

accuracy change over time. Also, there is an obvious

increase when the number of classes rises from 7 to 8.

These phenomena can explain the reason for the long

time cost of using SVM in this image classification

task.

5 CONCLUSION

For this image classification task, using SVM,

decision tree and the KNN algorithm can acquire a

precise result. These methods can be widely used in

other image classification tasks by combining with

other technology. If the algorithm listed in this paper

is combined them with the camera, it can be used for

the statistics of supermarkets. This will greatly

improve the operating efficiency of the supermarket

and reduce operating costs. However, there are still

some details that need to be improved. Due to the

limitations of those given data, future research needs

DAML 2023 - International Conference on Data Analysis and Machine Learning

372

to focus on enhancing algorithm accuracy. For

example, add an analysis of some of the features of the

image. By analyzing the texture of the image and the

shape of fruit, the accuracy of the classification

algorithm is improved.

REFERENCES

B. A. Yassine, B. Samia and O. Abdelaziz, An Improved

KNN Classifier for 3D Face Recognition Based on

SURF Descriptors. Journal of Applied Security

Research(4), 2023.

Y. Guo, Z. Tan and Y. Zhang, Glass component

classification model based on decision tree and cluster

analysis..(eds.)Proceedings of 2023 International

Conference on Mathematical Modeling, Algorithm and

Computer Simulation (MMACS 2023), 2023, pp. 221-

227.

Y. Zhang, S. Wang, G. Ji, P. Phillips, Fruit classification

using computer vision and feedforward neural network.

Journal of Food Engineering,2014, 143, pp. 167–177.

B. Wang, R. Ma, J. Kuang and Y. Zhang, "How Decisions

Are Made in Brains: Unpack “Black Box” of CNN

With Ms. Pac-Man Video Game," in IEEE Access, vol.

8, 2020, pp. 142446-142458.

M. A. Chandra and S. S. Bedi, Survey on SVM and their

application in image classification. International

Journal of Information Technology, 2018, 13(5), 1–11.

https://doi.org/10.1007/s41870-017-0080-1

J. Nie, K. Xu and Y. Chen, A glass classification model for

glass artifacts based on PCA and

SVM..(eds.)Proceedings of 2023 International

Conference on Mathematical Modeling, Algorithm and

Computer Simulation (MMACS 2023), 2023, pp. 286-

291.

P. Soucy and G. W. Mineau, A simple KNN algorithm for

text categorization. Proceedings 2001 IEEE

International Conference on Data Mining, 2002.

https://rachelchen0104.medium.com/hands-on-machine-

learning-with-scikit-learn-tensorflow-bbb1c91cb128

https://zhuanlan.zhihu.com/p/141472361

H. Li, An Overview on Remote Sensing Image

Classification Methods with a Focus on Support Vector

Machine.Ying Xing.(eds.)Proceedings of the 2021

International Conference on Signal Processing and

Machine Learning (CONF-SPML 2021), 2021, pp. 64-

70.

Fruit Image Classiﬁcation Based on SVM, Decision Tree and KNN

373