Journal of Physics: Conference Series

PAPER • OPEN ACCESS

X-DenseNet: Deep Learning for Garbage Classification Based on Visual

Images
To cite this article: Sha Meng et al 2020 J. Phys.: Conf. Ser. 1575 012139

View the article online for updates and enhancements.

This content was downloaded from IP address 193.31.72.18 on 14/07/2020 at 14:37

ISAI 2020 IOP Publishing
Journal of Physics: Conference Series 1575 (2020) 012139 doi:10.1088/1742-6596/1575/1/012139

X-DenseNet: Deep Learning for Garbage Classification Based

on Visual Images

Sha Meng, Ning Zhang and Yunwen Ren

Dalian Maritime University, 1th Linghai Road, Dalian City, Liaoning Province, China
Email: [email protected]

Abstract. In order to effectively solve the problem of garbage classification, this paper designs
a garbage classification model based on deep convolutional neural network. Based on Xception
network, combined with the idea of dense connections and multi-scale feature fusion in
DenseNet, the X-DenseNet is constructed to classify the garbage images obtained by visual
sensors. This paper conducts experiments through the process of “obtaining
dataset-preprocessing data-building X-DenseNet model-training and testing model” and the
accuracy of the model on the testing set is up to 94.1%, which exceeds some classic
classification networks. The X-DenseNet automatic garbage classification model based on
visual images proposed in this paper can effectively reduce manual investment and improve the
garbage recovery rate. It has the vital scientific significance and application value.

1. Introduction
With the development of economy, the total amount of garbage in the world is increasing year by year.
It is necessary to do garbage classification. As we all know, automatic classification which saves a lot
of manpower is becoming a development trend of garbage classification. However, the traditional
vision-based automatic garbage classification system uses simple manual features, and the
generalization performance is unsatisfactory. There are still many problems in the automatic
classification of complex and diverse garbage.
In recent years, deep learning has shown explosive development. Compared with traditional visual
characteristics extraction methods, the advantage of deep learning is that it does not need to select in
advance which features or design features to extract, but allows the model to learn from large-scale
data. So deep learning has a stronger learning ability and adaptability.
This paper proposes a visual sensor-based method, according to deep learning, in order to achieve
automatic garbage classification, which greatly improves the garbage collection rate.

2. Related Work
In the traditional method, Chan Xiang et al. [1] used Arduino UNO R3 as the main control board, used
the color recognition module to realize the classification function. Yet, in this method, the
characteristics of recognition are relatively simple, and the recognition object is relatively single. It is
difficult to perform efficient automatic classification of complex garbage.
In the deep learning method, Ning Kai et al. [2] used YOLOv2 as the main network module,
embedded deep dense modules to design a smart sweeping robot that can divide garbage into 25
subcategories including toilet paper, cans, milk, etc. according to shape and volume. Chen Yuchao et al.
[3] proposed to use OpenCV to implement the background difference algorithm, cut the object to be
detected from the picture, and then use the MobileNet to divide the image into syringes, hemostatic
forceps, infusion bags, and gloves.
Content from this work may be used under the terms of the Creative Commons Attribution 3.0 licence. Any further distribution
of this work must maintain attribution to the author(s) and the title of the work, journal citation and DOI.
Published under licence by IOP Publishing Ltd 1
ISAI 2020 IOP Publishing
Journal of Physics: Conference Series 1575 (2020) 012139 doi:10.1088/1742-6596/1575/1/012139

In summary, the garbage classification method based on deep learning has certain advantages.
Therefore, deep learning is of certain practical significance and scientific research value to solve the
problem of garbage classification and improve resource utilization.

3. Task Definition
This paper studies the problem of automatic classification of different types of garbage images in the
case of using a color camera to obtain color images in the actual garbage classification scenario. This
paper expects to generate the object category directly from the color image, and defines the function
M as the mapping relationship between the input image I and the output category C:
M I =C (1)
This paper uses a deep convolutional neural network to approximate the complex function M:
I→C. Let 𝑀𝜃 represent the neural network where 𝜃 is the weight of the network.
This paper proves that Mθ (I)=C≈M(I), you can use the training set input IT and the corresponding
output CT to learn and train the cross-entropy loss function L, as follows:
θ= min L CT ,Mθ IT (2)
θ

The practical application process is as follows:

 First, use a color camera to obtain a two-dimensional color image I0 =RH0 ×W0 with a height
of H0 and a width of W0 ;
 The obtained original image I0 is subjected to data preprocessing operations such as cropping
and denoising to obtain the processed image I1 =RH1 ×W1 ;
 Input the processed image I1 to the trained deep learning model Mθ to obtain the image
category output C=Mθ (I1 );
 Finally, according to the category C of garbage, the corresponding classification processing
operation is performed.

4. X-DenseNet Network
First of all, the Xception [4] network structure is used as the basic structure. With higher accuracy and
a relatively lower number of parameters, Xception is a further improvement of InceptionV3 [5]. The
added ResNet [6] residual network mode also makes the speed of convergence of the network
improve.
In order to make full use of image information features, this paper combines the idea of dense
connection and multi-scale feature fusion in DenseNet [7] in the network structure. Among them,
Dense Block is one of the main components of DenseNet. In a Dense Block, any layer is directly
connected to all subsequent layers. Therefore, the layer l receives the feature information of all
previous layers (x0 ,……,xl -1) as layer l input:
xl =Hl x0 ,x1 ,……,xl -1 (3)

Among them, [x0 ,x1 ,……,xl -1] refers to the connection of the feature map generated in the layer 0
to the layer l − 1.
Dense Block does not need to re-learn redundant parameter feature maps. In addition, each layer
can directly access the gradient from the loss function to the original input signal, which can
effectively slow down the problem of gradient disappearance. The structure of Dense Block adopted in
this paper is shown in Figure 1.

2
ISAI 2020 IOP Publishing
Journal of Physics: Conference Series 1575 (2020) 012139 doi:10.1088/1742-6596/1575/1/012139

Figure 1. The Structure of Dense Block

In order to merge multi-scale information and improve model performance, this paper on the basis
of Xception uses Dense Block to realize feature reuse and fusion. The X-DenseNet shown in Figure 2
is proposed. The specific model structure is as follows:
 Take the garbage image (150×150×3) as the input, first extract the features through the
Xception, and then transpose the convolution through 32 convolution kernels (3×3) and 64
convolution kernels (5×5) to increase the size and number of features, and then obtain feature
images 𝑥0 . The 𝑥0 is used as the input 𝐻1 of the lower layer Dense Block.
 In Dense Block, 𝐻1 undergoes batchnormalization and ReLu activation, and uses 32
convolution kernels (1×1) and 32 convolution kernels (3×3) for convolution, and finally obtains
feature images 𝑥1 . After splicing 𝑥0 and 𝑥1 , it is used as the input of the next layer 𝐻2 in
Dense Block, and so on. Between Dense Blocks, we use a convolution (1×1) and a maxpooling
(2×2) to connect Dense Blocks.
 The output of the third Dense Block is alternately transposed and convolved with 3×3 and 5×5
convolution kernels to expand the number of features. The final output passes through the fully
connected layer and finally maps to six neurons and then six types of garbage are classified.

Figure 2. The Structure of X-DenseNet

5. Experiment Preparation

5.1. Experimental Hardware Conditions

In this paper, a comparative experiment including X-DenseNet and AlexNet [8], ResNet50,
InceptionV3, Vgg16 [9], and Vgg19 [9] was performed on a remote server. Hardware devices that the
entire experimental process performed are shown in Table 1.
Table 1. Experimental Hardware Conditions
Version
Operating System Ubuntu 16.04.6 LTS 4.15.0-91-gerneric GNU/Linux
CPU Intel® Xeon(R) CPU E5-2670 0
GPU GeForceGTX1080Ti
PyCharm PyCharm 5.0.3
Deep Learning Framework Keras 2.3.1

3
ISAI 2020 IOP Publishing
Journal of Physics: Conference Series 1575 (2020) 012139 doi:10.1088/1742-6596/1575/1/012139

5.2. Dataset Generation

There are a total of 2527 color images [10] in the format of .jpg in the garbage image dataset used in
this paper, which are divided into 6 categories of garbage. There are 501 glass pictures, 594 paper
pictures, 403 cardboard pictures, 482 plastic pictures, 410 metal pictures and 137 other trashes pictures.
The size of these pictures is adjusted to 512 * 384, and the camera equipment is Apple iPhone. In
order to enhance the generalization performance of the model, the original image in the dataset is first
subjected to data enhancement operations through shear rotation, flipping and other methods. Finally,
90% and 10% of the pictures in the data set are used as the training set and testing set relatively.

6. Experimental Results

6.1. Training Results

This paper tested AlexNet, ResNet50, InceptionV3, Vgg16, Vgg19 and X-DenseNet in the experiment.
Each batch in the training process is 32. The optimizer is SGD, the momentum is 0.9, the initial
learning rate is 1e-3, and the loss function is the cross-entropy function.
The results of training accuracy and training loss are shown in Figure 3 and Figure 4. In terms of
convergence speed, AlexNet training convergence speed is very slow, the training accuracy reached 65%
in the 130th generation, and then stabilized at 70% afterwards. Vgg16 and Vgg19 have similar
convergence rates, reaching 95% and 93% respectively at the 100th generation. After 100 generations,
the training speed increases slowly. X-Dense and ResNet50 and InceptionV3 achieved 98% in training
accuracy in the 25th generation. It can be seen from the training accuracy of the 125 ~ 150 generations
that the training accuracy of X-Dense is almost always above 99%, and it is more stable than others.

Figure 3. Training Accuracy Figure 4. Training Loss

6.2. Testing Results

248 pictures in the data set were used as the test set. After data enhancement processing, each batch of
data in the test was 40 color garbage pictures. X-DenseNet is compared with AlexNet, ResNet50,
InceptionV3, Vgg16, and Vgg19, then the classification accuracy is got, as shown in Table 2.
X-DenseNet has the highest accuracy for the three types of garbage: cardboard, glass, and trash.
Compared with other models, X-DenseNet has a higher accuracy for identifying each type of garbage.
Overall, X-DenseNet's recognition efficiency is higher.
Table 2. The Testing Accuracy of Different Models
Cardborad/% Glass/% Metal/% Paper/% Plastic/% Trash/% Average/%
AlexNet 47.50 52.00 60.98 86.44 41.67 38.46 62.10
Vgg16 75.00 90.00 92.68 98.31 83.33 92.31 88.61
Vgg19 82.50 86.00 95.12 96.61 97.92 84.62 90.46
ResNet50 87.50 88.00 97.56 98.31 81.25 100.00 92.10
InceptionV3 85.00 90.00 92.68 96.61 79.17 100.00 90.58
X-DenseNet 95.00 90.00 95.12 94.92 89.58 100.00 94.10

6.3. Prediction Results

This paper takes 150 × 150 color garbage pictures as input, uses the X-DenseNet to classify and
predict the pictures, and obtains the prediction result pictures, as shown in Figure 5. In the results, the

4
ISAI 2020 IOP Publishing
Journal of Physics: Conference Series 1575 (2020) 012139 doi:10.1088/1742-6596/1575/1/012139

prediction and label are under each picture. The categories corresponding to each value in the list are
cardboard, glass, metal, and paper, plastic, and other trashes.

Figure 5. Prediction Results on X-DenseNet

7. Conclusion
Considering the problems of low efficiency of artificial garbage classification and poor generalization
performance of traditional garbage classification algorithms, this paper designs a garbage
classification model based on deep convolutional neural network X-DenseNet. The classification of
domestic garbage images obtained by visual sensors can effectively classify garbage automatically,
and divide the garbage into six categories: cardboard, glass, metal, paper, plastic and other garbage.
And the recognition accuracy rate is higher than other advanced image classification networks.

8. References
[1] Zhan Xiang, Xue Peng, Guo Huanping. Intelligent sorting trash can based on Arduino
[J].Electronic World, 2020 (04): 160-161.
[2] Ning Kai, Zhang Dongbo, Yin Feng, et al. Rubbish detection and classification of intelligent
sweeping robot based on visual perception [J]. Journal of Image and Graphics, 2019, 24 (8):
1358-1368.
[3] Chen Yuchao, Bian Xiaoxiao. Medical garbage classification system based on machine vision
and deep learning [J]. Computer Programming Skills and Maintenance, 2019, (5): 108-110. DOI:
10.3969 / j.issn.1006-4052.2019. 05.042.
[4] Francois Chollet. Xception: Deep Learning with Depthwise Separable Convolutions[C]// 2017
IEEE Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, 2017.

5
ISAI 2020 IOP Publishing
Journal of Physics: Conference Series 1575 (2020) 012139 doi:10.1088/1742-6596/1575/1/012139

[5] C. Szegedy, V. Vanhoucke, S. Ioffe, J. Shlens, and Z. Wojna. Rethinking the inception
architecture for computer vision. arXiv preprint arXiv:1512.00567, 2015.
[6] He K, Zhang X, Ren S, et al. Deep Residual Learning for Image Recognition[C]// IEEE
Conference on Computer Vision & Pattern Recognition. IEEE Computer Society, 2016.
[7] Huang G, Liu Z , Maaten L V D , et al. Densely Connected Convolutional Networks[C]// CVPR.
IEEE Computer Society, 2017.
[8] Krizhevsky A, Sutskever I , Hinton G . ImageNet Classification with Deep Convolutional
Neural Networks[C]// NIPS. Curran Associates Inc. 2012.
[9] Simonyan, Karen, Zisserman, Andrew. Very Deep Convolutional Networks for Large-Scale
Image Recognition[J]. Computer Science, 2014.
[10] garythung, “garythung/trashnet,” GitHub. [Online].
Available:https://github.com/garythung/trashnet.