Application of Deep Learning Based on Garbage Image Classification

FENG LI, LINGLING WANG*

School of Management Science and Engineering

Anhui University of Finance and Economics

Bengbu 233030, CHINA

Abstract: - In recent years, with the rapid development of economy, the country's various construction is

thriving, and remarkable achievements. At the same time, resources and the environment have been seriously

damaged. This phenomenon is directly related to the irrationality of garbage classification and delivery, and the

contradiction between the two is becoming increasingly acute as people strongly reflect the problem of

environmental pollution but do nothing about it. This paper designs a garbage image classification system

based on deep learning, the main research content is to compare multiple deep learning neural network models,

find the optimal classifier, develop web applications and deploy neural networks, which includes image data

acquisition, image pre-processing, and comparison of VGG16, Inception, and Resnet neural network model

accuracy

Key-Words: Garbage Image Classification; Deep Learning; VGG16; Inception.

Received: September 19, 2021. Revised: May 17, 2022. Accepted: June 13, 2022. Published: July 15, 2022.

1 Introduction

At present, the main way of garbage

classification is that residents consciously classify

garbage when they dump it and throw it into the

corresponding garbage [1-3]. However, this garbage

classification method based on residents'

consciousness is difficult to be deeply implemented,

and residents cannot accurately distinguish the types

of recyclable garbage and classify each kind of

garbage in detail [4]. Through some city residents

forced garbage classification in recent years, but

through a period of implementation, the effect is not

good, most of the general public is not entirely

correct in classifying rubbish, so to enforce garbage

classification was fined not only let the masses

themselves, and also for the recyclable garbage

classification is not too big effect [5].

Therefore, it is not an effective method to

classify garbage solely by the waste producers'

consciousness. Instead, it should be simply

classified by the waste producers, then the waste

recycling plant will carry out detailed garbage

classification and classify and treat the recyclable

garbage [6]. Garbage problem is increasingly

serious, with the rapid development of economy, the

urbanization process is gradually accelerating,

people's living water. In order to meet people's

growing needs for life, a variety of diversified

commodities and articles of daily use have been

pouring in. This is followed by an increasing

amount of household garbage, which has caused

great pressure on the environment. Some cities have

already seen garbage [7]. The phenomenon of siege,

garbage classification is a reform of the traditional

way of garbage collection and disposal, is an

effective disposal of garbage management methods.

How to deal with recyclable waste is a major

research issue globally. The classification and reuse

of recyclable garbage is an important link between

resource recycling and reuse [8]. The effect of

garbage resource recycling is closely related to the

quality of garbage classification. Rapid and

effective garbage classification can greatly reduce

the environmental pollution and resource waste

caused by garbage. The classification of recyclable

garbage by deep learning has the characteristics of

high precision, fast speed and high adaptability [9]

[10]. The establishment of a precise garbage

WSEAS TRANSACTIONS on COMPUTERS

DOI: 10.37394/23205.2022.21.34

Feng Li, Lingling Wang

E-ISSN: 2224-2872

277

Volume 21, 2022

classification system can not only help garbage

recycling plants to carry out garbage classification,

because of its simple operation, but also help people

to identify the types of recyclable garbage, help

them to accurately put garbage, and realize the

intelligent development of garbage classification.

An automated machine that can automatically

sort recyclable garbage [11]. The machines input

images from external devices, such as cameras, and

then calculate the types of garbage using a deep

learning network inside the machine. At the time,

however, the machine could only sort the valuable

rubbish from the huge amount of rubbish waiting to

be sorted, not really sort the huge amount of rubbish.

Subsequently, computer vision technology has been

widely applied to image detection and classification

in the academic world, and computer vision is used

to classify various objects in the industry. However,

due to the diversity of recyclable garbage, mature

garbage classification algorithm is still a difficult

problem. Plastic accounts for a large proportion of

recyclable waste, and the accuracy of a traditional

computer vision-based algorithm for classifying

plastic bottles is greatly reduced due to the

transparent nature of plastic and the complex

background of garbage classification [12]. Later, a

plastic sorting method based on infrared spectrum

analysis equipment was also proposed, but this

equipment is relatively expensive, so it is not

suitable for large-scale commercial use. It can be

seen that the traditional computer vision is not ideal

for the classification of recyclable garbage. On the

one hand, the identification accuracy is not high

enough; on the other hand, the classification cost is

too high, which is not suitable for large-scale

application in manufacturers.

2 Overview of Deep Learning

Deep Learning is a kind of neural network

technology. Its most revolutionary point is that as

long as there is enough learning data, the neural

network itself can automatically extract the features

of the data group. The analysis of images and data

before this requires the operation of the extraction

algorithm according to each data and problem.

However, deep learning does not require manual

operation, but automatically extracts features.

Slightly crudely, this is what it means: simply inject

data into a neural network and you can extract any

feature you want.

A neural network is a network of connected

brain nerve units that mimic neurons [13]. The input

signal is propagated. One of the characteristics of

this method is that each layer of neural network can

be phased learning. For example, let the first layer

learn to output the input information as is, let the

second layer learn to reproduce the input in the

same way on the basis of the first layer, the third

layer after the same operation. Deep learning is best

at pattern recognition of data that cannot form

symbols, such as image data and waveform data,

and stage learning after inputting images through

the input layer. The commonly used neural network

is a perceptual neural network with all layers

connected together. But in the case of image

recognition, it is easier to adopt a special connection

method. This is called a convolutional neural

network. It is the prototype of deep learning. It is

characterized by a multi-scale intermediate layer for

different size segmentation of input data and feature

extraction. Input an image of a car, for example, and

you can extract everything from detailed patterns to

large structures and overall Outlines.

Deep learning is a division in the field of

artificial intelligence as a whole. Some researchers

believe that the neural network before deep learning

has made a leap forward, but there is still some gap

between it and the social cognition that suddenly

approaches human beings [14]. Moreover, it takes

time and stages to reach a point where it is truly

useful to society. At present, the development of

technology for practical purposes in many fields is

proceeding rapidly all over the world. In particular,

the neural network is characterized by the repetition

of market calculation and the large number of

parallel calculations. As a result, graphics boards

equipped with gpus, which support the gaming and

computer graphics industries can be used.

In a word, deep learning can help us classify

recyclable garbage reasonably and efficiently, and

create a society in which garbage resources can be

reasonably reused and resources can be fully

WSEAS TRANSACTIONS on COMPUTERS

DOI: 10.37394/23205.2022.21.34

Feng Li, Lingling Wang

E-ISSN: 2224-2872

278

Volume 21, 2022

recycled. Deep learning technology has a mature

application case in the field of classification of

various images, so it is feasible to use neural

network for garbage classification technically. At

the same time, the classification accuracy of garbage

image is low by using traditional image processing

methods, but the classification of garbage image by

using deep learning can quickly improve the speed

and accuracy of garbage image classification.

3 Proposed Method

Convolutional neural network is an important

component of current deep learning [15].The

emergence of convolutional neural network and

back propagation algorithm makes deep learning

and artificial intelligence enter a new stage of

development. Convolutional neural network is

designed after the structure of biological neurons. It

is generally divided into convolution layer, pooling

layer, full connection layer, activation function, loss

function and random inactivation.

Convolutional neural network can extract

features through the sliding window in the

convolutional layer, and then summarize the

extracted features through the full connection layer,

which accounts for a large proportion of the

parameters of the whole neural network. Then the

pooling layer is responsible for compressing and

simplifying the matrix of the output of the previous

hidden layer, so as to facilitate feature sampling.

Finally, the output of convolutional neural network

similar to image classification cannot be directly

utilized by us, but the output of the model should be

transformed into various categories through the

model structure similar to classifier.

The loss function, the activation function and

the random inactivation is a part of the model

control effect, loss function is the key to the

convolution back propagation neural network

model, through calculating the loss function, the

model can be error between the predicted value and

actual value, which based on loss function to adjust

the parameters in the model. Activation function is

to prevent only linear changes in the whole model.

The whole network can not change in a linear

direction through activation function. Typical

activation functions include Sigmod function and

Relu function. Random inactivation function will

randomly discard some parameters, so that the

whole network will not overfit in the training

process..

After obtaining the data that can be used for

testing, three models VGG16[16], Xception[17] and

InceptionResnetv2 are respectively used for transfer

learning in order to obtain better detection effect,

and the optimal model is selected by comparison.

Finally, InceptionResnetv2 was selected as the

model according to the average accuracy and

prediction speed of the model.

In order to extract image features more

effectively, Inception module uses a structure that

combines several different convolution operations in

parallel to extract image features more effectively.

But the corresponding tighter module also makes

the parameters of the module larger. At the same

time, the Resnet structure can not only accelerate

the training speed, but also simplify the complexity

of the model and prevent the gradient diffusion.

The author of the model combined Resnet and

Inception structures, so that the model would not

have a large increase in the number of parameters

without reducing the accuracy of the model,

resulting in high training costs. Figure 1 shows the

model structure.

This model is continuously iterated from

Inception series models, and combined with Resnet

residual network, a downsampling structure is added

to increase model parameters so as not to cause

explosive growth of model parameters. The model

starts with a 299×299 pixel RGB image input

through the Stem structure of the first layer, which

includes multiple convolution processes and merge

operations. The matrix data through Stem will enter

the following multiple Inception-resnet structures.

The general principle of these structures is the same.

They are all carried out in parallel by multiple

convolution operations, and then merged. At the

same time, the Inception-resnet module is different

from the previous Inception module in that it adopts

the residual structure, which greatly reduces the

WSEAS TRANSACTIONS on COMPUTERS

DOI: 10.37394/23205.2022.21.34

Feng Li, Lingling Wang

E-ISSN: 2224-2872

279

Volume 21, 2022

Fig.1. Overviewof InceptionResnetv2

time required for training the model and achieves

higher accuracy in fewer epoches.

In the final Average Pooling operation, the

Average Pooling structure carries out the Average

sampling operation on the model of the upper layer,

which replaces the fully connected layer and greatly

reduces the number of parameters of the model.

Finally, the one-dimensional parameters after

average pooling are transferred to the linear

classification layer of the network through Dropout

processing. Finally, Softmax function is used to

normalize the obtained parameters, and finally the

probability values of different types of model

prediction are obtained.

4 Experimental

In this section, we firstly collect the recyclable

garbage data, and then data preprocessing is to

convert the original obtained data into the data

format that meets the requirements of the project

through a series of transformations.

4.1 Data Collection

In order to obtain the common recyclable

garbage data, and the garbage data acquisition and

pretreatment, in order to obtain qualified data set, so

as to carry out the training and construction of

identification system. The main collection methods

are as follows:

(1) Crawler acquisition: Requset + Selenium is

used to obtain image data from large search

websites such as Baidu by directly searching target

items and crawling them with Python crawlers. The

data crawled by crawler are generally messy and

need to be manually cleaned to delete pictures of

non-target objects and repeated pictures.

(2) Related projects: Transplant the data used

in part of the old garbage identification system.

Clean data, do not need too much cleaning, storage

by category.

(3) Data filling: For the above two methods to

obtain a sufficient number of garbage categories, we

use to search some well-known image online and

second crawler to fill the database imbalance.

After obtaining the summary data, we obtained

a total of 17,088 pictures, including 24 categories,

respectively: Pans, cutting boards, leather shoes,

bottle, bag, pillow, beverage bottle, socket, old

clothes, cans, charging treasure, cardboard boxes,

glasses, condiment bottles, plastic bowl bowl,

plastic hangers, plastic toys, plug wire, shampoo

bottle, plush toys, cosmetics bottles, Courier bags,

metal food cans, food cans, are common in daily life

of recycled garbage.

4.2 Data Processing

Data preprocessing is to convert the original

obtained data into the data format that meets the

requirements of the project through a series of

transformations. Whether the data format is correct

or not has a crucial impact on the accuracy of model

WSEAS TRANSACTIONS on COMPUTERS

DOI: 10.37394/23205.2022.21.34

Feng Li, Lingling Wang

E-ISSN: 2224-2872

280

Volume 21, 2022

training. After obtaining 24 categories of data, 24

folders are read respectively, and a series of

operations are carried out on the data in them to

produce data sets that meet the requirements of the

model.

(1) Image renaming. Through traversing each

image under 24 folders, all the images are stored in

the format of category name + serial number, and all

the serial numbers are stored in the format of five

digits (all the insufficient digits are filled with 0).

Data preprocessing is to convert the original

obtained data into the data format that meets the

requirements of the project through a series of

transformations. Whether the data format is correct

or not has a crucial impact on the accuracy of model

training. After obtaining 24 categories of data, 24

folders are read respectively, and a series of

operations are carried out on the data in them to

produce data sets that meet the requirements of the

model.

(2) Image renaming. Through traversing each

image under 24 folders, all the images are stored in

the format of category name + serial number, and all

the serial numbers are stored in the format of five

digits (all the insufficient digits are filled with 0).

4.3 Experiment results

The training results of the model need certain

evaluation standards. Firstly, a new validation data

set should be created in addition to the training data

set of the model to verify the accuracy and loss

value of the model. When verifying the accuracy of

the model, the data set during the training of the

model cannot be used, because the model may have

over-fitting, that is, the model has a very high fitting

degree to the training data, but the prediction

accuracy of the data other than the training data

used is very low, or even cannot be fitted basically.

Therefore, a separate validation dataset needs

to be created to calculate the exact results of the

model. InceptionResnetv2 has a detection accuracy

of 89% and a loss value of 0.8. Compared with the

above two migration models, InceptionResnetv2 has

better detection performance, as shown in Fig.2.

Fig. 2. Experiment results

5 Conclusions

In this paper, we design a garbage image

classification system based on deep learning

method, the main research content is to compare

multiple deep learning neural network models, find

the optimal classifier, develop web applications and

deploy neural networks, which includes image data

acquisition, image pre-processing, and comparison

of VGG16, Inception, and Resnet neural network

model accuracy

Acknowledgment

We thank the anonymous reviewers and editors

for their very constructive comments. This work

was supported in part by the Natural Science

Foundation of the Higher Education Institutions of

Anhui Province under Grant No. KJ2020A0011,

WSEAS TRANSACTIONS on COMPUTERS

DOI: 10.37394/23205.2022.21.34

Feng Li, Lingling Wang

E-ISSN: 2224-2872

281

Volume 21, 2022

Innovation Support Program for Returned Overseas

Students in Anhui Province under Grant No.

2021LCX032. the Science Research Project of

Anhui University of Finance and Economics under

Grant No. ACKYC20085, Undergraduate teaching

quality and teaching reform project of Anhui

University of Finance and Economics under Grant

No. acszjyyb2021035.

References:

[1] Szaky, Tom. Outsmart waste: The modern idea of

garbage and how to think our way out of it. Berrett-

Koehler Publishers, 2014.

[2] Anantharaman, Manisha. "Networked ecological

citizenship, the new middle classes and the

provisioning of sustainable waste management in

Bangalore, India." Journal of Cleaner Production 63

(2014): 173-183.

[3] Okwesili, Joseph, and Chidi Iroko. "Urban solid

waste management and environmental sustainability

in Abakaliki Urban, Nigeria." European Scientific

Journal 12.23 (2016).

[4] Tai, Jun, et al. "Municipal solid waste source-

separated collection in China: A comparative

analysis." Waste management 31.8 (2011): 1673-

1682.

[5] Meng, Xiaoyan, et al. "Investigation on decision-

making mechanism of residents’ household solid

waste classification and recycling behaviors."

Resources, Conservation and Recycling 140 (2019):

224-234.

[6] Gille, Zsuzsa. From the cult of waste to the trash

heap of history: the politics of waste in socialist and

postsocialist Hungary. Indiana University Press,

2007.

[7] Junk, Wolfgang J., et al. "Current state of

knowledge regarding the world’s wetlands and their

future under global climate change: a synthesis."

Aquatic sciences 75.1 (2013): 151-167.

[8] Singh, Jagdeep, and Isabel Ordoñez. "Resource

recovery from post-consumer waste: important

lessons for the upcoming circular economy." Journal

of Cleaner Production 134 (2016): 342-353.

[9] Bircanoğlu, Cenk, et al. "RecycleNet: Intelligent

waste sorting using deep neural networks." 2018

Innovations in intelligent systems and applications

(INISTA). IEEE, 2018.

[10] Sudha, S., et al. "An automatic classification method

for environment: Friendly waste segregation using

deep learning." 2016 IEEE Technological

Innovations in ICT for Agriculture and Rural

Development (TIAR). IEEE, 2016.

[11] Gundupalli, Sathish Paulraj, Subrata Hait, and Atul

Thakur. "A review on automated sorting of source-

separated municipal solid waste for recycling."

Waste management 60 (2017): 56-74.

[12] Villanueva, Alejandro, and Peter Eder. "End-of-

waste criteria for waste plastic for conversion."

Institute for Prospective Technological Studies

(2014).

[13] Neelakanta, Perambur S., and Dolores F. De Groff.

Neural network modeling: Statistical mechanics and

cybernetic perspectives. CRC Press, 2018.

[14] Mascagni, Pietro, et al. "Artificial intelligence for

surgical safety: automatic assessment of the critical

view of safety in laparoscopic cholecystectomy

using deep learning." Annals of surgery 275.5

(2022): 955-961.

[15] Bashar, Abul. "Survey on evolving deep learning

neural network architectures." Journal of Artificial

Intelligence 1.02 (2019): 73-82.

[16] Simonyan, Karen, and Andrew Zisserman. "Very

deep convolutional networks for large-scale image

recognition." arXiv preprint arXiv:1409.1556

(2014).

[17] Chollet, François. "Xception: Deep learning with

depthwise separable convolutions." Proceedings of

the IEEE conference on computer vision and pattern

recognition. 2017.

Sources of funding for research

presented in a scientific article or

scientific article itself

Creative Commons Attribution License 4.0

(Attribution 4.0 International, CC BY 4.0)

This article is published under the terms of the Creative

Commons Attribution License 4.0

https://creativecommons.org/licenses/by/4.0/deed.en_US

WSEAS TRANSACTIONS on COMPUTERS

DOI: 10.37394/23205.2022.21.34

Feng Li, Lingling Wang

E-ISSN: 2224-2872

282

Volume 21, 2022