A SMART RECYCLING BIN USING WASTE IMAGE

CLASSIFICATION AT THE EDGE

Preprint, compiled October 4, 2022

Xueying Li and Ryan Grammenos

[email protected] , [email protected]
Electrical and Electronic Engineering
University College London

Abstract
arXiv:2210.00448v1 [cs.CV] 2 Oct 2022

Rapid economic growth gives rise to the urgent demand for a more efficient waste recycling system.
This work thereby developed an innovative recycling bin that automatically separates urban waste
to increase the recycling rate. We collected 1800 recycling waste images and combined them with
an existing public dataset to train classification models for two embedded systems, Jetson Nano and
K210, targeting different markets. The model reached an accuracy of 95.98% on Jetson Nano and
96.64% on K210. A bin program was designed to collect feedback from users. On Jetson Nano, the
overall power consumption of the application was reduced by 30% from the previous work to 4.7
W, while the second system, K210, only needed 0.89 W of power to operate. In summary, our work
demonstrated a fully functional prototype of an energy-saving, high-accuracy smart recycling bin,
which can be commercialized in the future to improve urban waste recycling.

I. Introduction replace the traditional trash bins at the locations without plug
sockets. Therefore, it can hardly increase the household and
Waste generation has increased dramatically in the 21st century street recycling rate. As a result, this paper proposes low-cost
due to the growth in the global population. According to battery-powered real-time recycling waste segmentation bin
the World Bank Group’s report [1], 2.01 billion tonnes of systems to fill this gap.
municipal solid waste are generated worldwide every year, In this research, we trained a light MobileNet image classifica-
with only about 19% adequately recycled [1]. Recycling not tion model for Jetson Nano with TrashNet [3], an open-source
only helps to preserve raw material but, more importantly, waste dataset, and 1800 new training samples collected by
reduces the landfills required, which is an undesirable way us. The model demonstrated great performance with a high
of waste disposal due to its high demand for space and the test accuracy of 95.98% and a small parameter size of 3.0 M.
danger of introducing contaminants into the ground or even Jetson Nano only consumed 4.7 W when it ran the model. We
groundwater system. A major step of recycling is to separate also found a cheaper and less energy-hungry device, K210
the waste into specific categories according to its material. [4] to further reduce power consumption. The K210 board
Failure to do so will significantly harm the effectiveness of only consumed 0.89 W at inference time and the model on
recycling. Traditionally, workers at the recycling company sort it achieved a high test accuracy of 96.64%, which made it a
the waste into corresponding categories by hand, which is an more practical solution for trash bin applications. The code
inefficient method and requires unnecessary labor. Therefore, used in this study is available on Github [5].
people are now in dire need of a more advanced and automated
waste separation system. This paper is organized as follows. In Section II, the related
works performed in waste classification and AI trash bins
Image classification machine learning (ML) algorithms have are introduced. Section III introduces the system design and
been used to build automatic waste classification systems and explains the background theories that support this study.
assist waste management. This research aimed to improve the The methodology of data collection and model training is
existing classification system and pursue the commercialization introduced in section IV. Section V analyses the test results
of an artificial intelligence (AI) street bin. Specifically, we and evaluates the system performance on Jetson Nano and
focused on reducing the power consumption of the controller K210. Finally, Section VI concludes the achievements and
board to extend AI bin’s battery life, increasing its waste provides suggestions for future research.
classification accuracy and reducing its price.
The only intelligent recycling bin on the market, Bin-e [2],
uses image classification algorithms to separate the trash into
II. Related works
four categories: plastic, paper, metal and glass. It achieves
a waste segmentation accuracy of 92% [2]. However, the The United Kingdom (UK) government has planned to increase
price USD 5200 [2] is expensive for a rubbish bin. Also, the household recycling rate in England to 50% by 2020,
Bin-e’s operation requires a 230V power supply, so it cannot but only 44% of municipal waste was reused and recycled
in 2020 [6]. The inconvenience and lack of knowledge are
Preprint – A SMART RECYCLING BIN USING WASTE IMAGE CLASSIFICATION AT THE EDGE 2

two crucial factors that prevent people from recycling [7]. The classification speed is quantified by inference time, which
Automatic waste segmentation bins were designed to overcome refers to the time taken for one classification to be completed.
these problems by helping people classify and send the waste Our application looks for a smaller model that can interact
into the corresponding containers, making waste disposal more with the user in real-time, with an inference time that must be
convenient. Table I outlines the approaches and the primary smaller than 0.1 s, which is the human visual reaction time. The
hardware components involved in related studies to construct CNN model, EfficientNet B0, used in [17] is the starting point
the waste segmentation systems. of this research and it achieved an accuracy of 95.38% and
an inference time of 0.07 s on Jetson Nano. While the model
In early research, a microcontroller is connected to a variety of
is fast enough to be used in real-time applications, the 96%
sensors to determine the composition of the waste. For example,
high memory usage needs optimization for bin applications.
inductive and capacitive sensors can detect the metal element
[8][9][10] and moisture sensors separate wet waste from dry Power consumption is another factor that needs to be consid-
waste [9][10]. The microcontroller makes decisions based ered in the final product. Unfortunately, rare research has paid
on its readings [8][9]. However, although the sensor-based attention to it. For instance, none of the trash bin cases listed in
classification method can detect the composition precisely, it Table I has measured the power consumption of the proposed
lacks the ability to classify waste into more specific groups. The system. Still, it is evident that the CNN-based approaches have
sensors cannot discriminate between plastic, paper, glass, and higher power consumption than the sensor-based one and the
other unrecyclable dry waste, which are important categories Pynq-Zl and Raspberry Pi will typically have lower power
in recycling. consumption than Jetson Nano used in [12] and [17]. The
Raspberry Pi 4 has a typical power consumption of between
The development of machine learning and image classification
2.7 W and 6.4 W. It reduces the power consumption but
enables the bin to sort the waste based on visual input
results in undesirable low performance [14]. As a result, the
like a human. The convolutional neural network (CNN) is
CNN architecture, MobileNet V2, has a low average per-class
a branch of image classification algorithms that performs
precision of 91.76% and a long inference time of 0.358 s on
mainly convolution operations on the pixels [15]. It is the most
it.
popular choice due to its high accuracy and power efficiency
compared to other methods and is used in all the four papers The limitation of the current research is that the energy-
[11][12][13][14]. It gives the bin ability to differentiate between saving systems will have unacceptable low performance in
spoons and cups, which is tremendous progress compared to the the classification task for commercialization. As a result, this
sensor-based approaches. Traditionally, the CNN models run on study designed two high-accuracy waste classification systems
a cloud raising the data transmission latency and user privacy with lower power consumption than previous studies. The first
security problems. To solve these problems, recent research system reduced the power consumption of the application on
moved the computation to an edge embedded system. However, Jetson Nano by using a lighter model, MobileNet, to reduce
edge computing has the drawback of limiting computation power consumption while maintaining the accuracy.
resources, so the model size is important in selecting the CNN
The second system is developed based on K210, a less
model structure.
expensive and more energy-saving embedded device. K210 is
To develop, evaluate and select the proper CNN structures, most an unpopular choice and is used in fewer than 100 research.
research in this field used the TrashNet dataset developed by Most research implemented YOLO object detection models on
Mindy Yang and Gary Thung [3] in 2017, the first open-access it [18][19] and demonstrated its outstanding power efficiency.
recycling waste dataset. This high-quality dataset contains K210 has not been used for recycling waste classification when
2527 waste photos from 6 groups: metal, glass, plastic, paper, this paper was written, so this paper proposed and evaluated
cardboard, and trash. It provides a foundation for later research an innovative approach.
and our study also used this dataset.
Efforts have been made to increase the segmentation accuracy III. System design
of CNN models based on TrashNet. The state-of-art accuracy
of 97.86% on TrashNet was reached in [16]. However, the A. The AI recycling bin design
CNN architecture, GoogLeNet, used in this research will cause
out-of-memory (OOM) on Jetson Nano [17]. It demonstrated The AI bin consists of five recycling waste containers and a
the potential of CNN classification, but the model size and detection box. The waste will be sorted after it is placed in
computation cost must be cut down to implement the machine the detection box. The whole system can be controlled by a
learning algorithm on embedded systems. A lighter CNN Jetson Nano or a K210 board.
model, WasteNet, was constructed by White et al. and achieved
an accuracy of 97% on the TrashNet dataset [12]. The paper The bin design using Jetson Nano is summarized in Figure
claimed that the model could be loaded to Jetson Nano, but 1. Jetson Nano will interact with the users and collects
it did not provide details regarding the edge implementation, feedback through the touch screen in front of it, displaying
such as the classification speed. Nevertheless, from the model the instructions for using the bin and the camera inputs. The
structure, we could estimate that the classification speed for Raspberry Pi camera at the top takes photos of the waste in
one image would be too slow for real-time classification. It the detection box. The images will be fed to the classification
can only be used in a bin application that classifies the waste algorithm in Jetson Nano, which classifies the photos into
objects based on one photo. seven groups. The waste classification models in the previous
study [17] have five output classes: "paper", "metal", "plastic",
Preprint – A SMART RECYCLING BIN USING WASTE IMAGE CLASSIFICATION AT THE EDGE 3

TABLE I
Related works in AI bins

Year Author Waste classification method Classification category The device

2014 Rajkamal et al. Inductive metal sensor, metal/glass, food, bio, PIC18F4580
[8] capacitive based moisture sen- paper/plastic, inert controller
sor, methane sensor, odor sen-
sor
2017 kumar et al. [9] Inductive and capacitive metal Biodegradable, plastic, metal STM32 controller
sensor, gas sensor, bacteria sen-
sor, moisture sensor
2019 Pereira et al. Infrared radiation sensor, capac- Dry, wet, plastic Atmega 328P
[10] itive sensor microcontroller
2019 Ziouzios et al. CNN image classfication Glass, paper, metal, plastic, Xilinx Pynq-Zl FPGA
[11] cardboard, trash
2020 White et al. CNN image classification Paper, plastic, metal, glass, Jetson Nano
[12] cardboard, other
2021 Jimeno et al. CNN image classification Aluminum cans, plastic Computer with
[13] bottles, plastic cups, paper, Nvidia RTX 2060 GPU
spoons/forks
2021 Sallang et al. CNN image classification Glass, paper, metal, plastic, Raspberry Pi 4
[14] cardboard

TABLE II
Training parameters used in [17]
Pre-Processing Method Value
Random flip Horizontal and vertical
Random rotation Up to 180◦
Random translation Up to 10%
Random zoom Up to 75%
Training Parameter Value
Learning rate scheduler Constant learning rate scheduler
Fig. 1. System block design of WasteNet. Train/Validation/Test split ratio 72/18/10
Optimizer Adam optimiser
Training epochs 50
"cardboard" and "glass". They will be reproduced in the next Learning rate 4.3e-05
section as the benchmark models of our study. Our model will Fine-tuning training epochs 8
be trained with two new classes, "empty" and "hand". "Empty" Fine-tuning learning rate 4e-06
Loss functions Sparse categorical cross entropy
means there is nothing in the photo, while the "hand" group
Classifier activation function Softmax
detects a human’s hand to avoid trapping the user’s hand by Include top layers in base model False
the door. When these two groups are detected, the bin waits Batch size 16
and continues detecting.

B. The Benchmark models and validation, and 238 images were used for testing. The
train/validation/test split is 72/18/10.
The waste classification models in [17] consist of an input
layer, a pre-processing augmentation layer, an EfficientNet B0 To begin with, the base model is initialized with pre-trained
base model layer, a global average pooling 2D layer and a weights on ImageNet. The two models are trained separately
dense layer. The first model has an input layer size of 512x384 by setting the corresponding input sizes. The learning rate
pixels, while the second model input has 384x288 pixels. Table was set to 4.3e-05 at the first 50 epochs and was reduced to
II shows the configuration of the augmentation layer and the 4e-06 at the eight epochs afterwards. Eventually, both models
model training hyperparameters for reproduction. achieved the same test accuracy of 95.38%. Figure 2 shows
the confusion matrix of the test results and indicates that most
The models were trained with the TrashNet dataset, which
mistakes are made in classifying between paper and cardboard.
consists of six categories of images: glass, paper, cardboard,
plastic, metal, and trash. Only the first five groups of data
were used to build the model that classified recycling into
five categories. In total, 2152 images were used for training
Preprint – A SMART RECYCLING BIN USING WASTE IMAGE CLASSIFICATION AT THE EDGE 4

TABLE III
Model classification accuracy on ImageNet
CNN architecture Number of Top-1 accuracy on
parameters (M) ImageNet (%)
EfficientNet B0 [23] 5.3 76.3
MobileNet V3 Large 5.4 75.2
[22]
MobileNet V1 [20] 4.2 70.6
ShuffleNet V2 [24] 7.4 74.9
Dense-169 [25] 14.3 76.2
ResNet50 [26] 25. 6 74.9

D. Jetson Nano and Kendryte K210 chip

The model is deployed to Jetson Nano in the previous study

[17]. Jetson Nano [27] is a powerful single-board computer
designed by Nvidia for AI applications. It runs the Linux
embedded system, which can execute multiple applications and
graphics processing unit (GPU) accelerated neural networks in
Fig. 2. Confusion matrix of the benchmark models parallel. Its developer kit has a high-definition multimedia
interface (HDMI) port, two camera serial interface (CSI)
camera slots and four universal serial bus (UCB) ports that can
After the model was reproduced, it was saved into TensorFlow connect to accessories. The 40-pin header allows it to drive
SavedModel format and transferred to Jetson Nano. The motors and receive signals from sensors. It has a 5 W mode
384x288 model was optimized successfully through TensorRT and a maximum power mode with typical power consumption
on Jetson Nano, but the 512x384 model caused an OOM error. between 5 W and 10W.
The throughput of the accelerated model was 19 inferences
per second (IPS) at the maximum power mode. In this study, we also explored the use of a K210 developer kit
to run the waste classification model. K210, in contrast, only
Nevertheless, the 384x288 model occupied 94.9% of memory. runs one model stored in its flash memory during operation
The system would freeze up if we ran other applications without an operating system. Canaan developed it to perform
simultaneously. To implement the classification model in an tasks specialized in image and audio processing. It has a neural
AI bin application, it is essential to replace the original base network processor called knowledge processor unit (KPU) used
model with the lighter model, such as MobileNet V3, which to accelerate CNN computations, such as convolution, batch
has lower memory usage and power consumption. normalization, activation, and pooling operations. Images can
be read from a digital video port (DVP) camera and displayed
C. MobileNet V3 model on a liquid-crystal display (LCD) screen connected to board.
The Sipeed M1w dock K210 developer kit has a maximum
The EfficientNet architecture used in the benchmark model power consumption of 3 W.
has demonstrated outstanding performance on the ImageNet
dataset. In particular, EfficientNet B7 has the same top- Table IV shows the specifications of the two embedded devices.
1 accuracy, 84.3%, as the GPipe while being 8.4x [20]. Jetson Nano has greater flexibility and computation power
However, as EfficientNet B0 is the smallest model in the for applications, while K210 has lower price and power
family, with a parameter size of 4.0 M without top fully consumption. K210 only supports 1x1 and 3x3 kernels in
connected layers, so MobileNet V3 Large, which only has 3.0M CNN, so it cannot implement most CNN structures such as
parameters and top-1 accuracy of 1.1% less than EfficientNet MobileNet V3. The low random-access memory (RAM) limits
on ImageNet, is chosen as the substitute. MobileNet is a the model size to 6 MB and the input resolution to 320x240
series of CNN architectures developed by Google. It contains pixels. In contrast, the 4GB RAM allows Jetson Nano to run
three architectures: MobileNet V1 [20], V2 [21] and V3 [22]. more complex models such as Efficient Net B0 with an input
MobileNet V2 achieved higher accuracy on ImageNet with a of 512x384 pixels. Therefore, K210 is a cheaper substitute for
smaller parameter size than MobileNet V1. The MobileNet Jetson Nano which runs models lighter than MobileNet V1.
V3 optimized the latency of MobileNet V2.
Other CNN models also achieved similar accuracies as
benchmark models, such as ShuffleNet V2, ResNet50 and
Dense-169 as listed in Table III, but their parameter size is
too large and results in high latency. The parameter size of
the MobileNet V3 Large is smaller than EfficientNet B0 when
the fully connected layers are removed in both models.
Preprint – A SMART RECYCLING BIN USING WASTE IMAGE CLASSIFICATION AT THE EDGE 5

TABLE IV
Comparison of Jetson Nano and K210
Nvidia Jetson Sipeed M1w dock
Nano 4GB[27] K210 kit[28]
Clock 1.43GHz 0.4GHz
RAM 4GB 8 MB
Power consumption 5-10W 0.3W-3W (typically
1W)
Programming language Python Micropython
Cost GBP 133 GBP 23

Fig. 3. A sample waste image taken by the cameras

After the models are trained in TensorFlow, they are optimized
for on-device machine learning before deployment to reduce
the size and latency. TensorFlow Lite library [29] can compress The "hand" group contains photos of the waste held by hand.
the TensorFlow model through quantization, which reduces the Figure 4 shows an example of the "hand" class. If there is a
precision of the weights from 32-bit floats to 16-bit floats or hand in the images, it belongs to the "hand" group, regardless
8-bit integers. K210 uses nncase [30], an open-source library of the type of waste present.
developed by Kendryte, to accelerate the TensorFlow Lite
model into its unique kmodel format that optimizes KPU usage
through constant-folding, operator replacement and operator
fusion. Firstly, constant-folding computes the expressions that
only contain known constants and substitutes the answers to
save computation cost at runtime. Then, operator replacement
changes some operators in the TensorFlow Lite model, which
KPU cannot accelerate, with several operators that perform a
similar task. Finally, operator fusion combines several operators
into a single KPU accelerated operator. These operations
optimizes the latency of model running on K210.
For Jetson Nano, Nvidia’s TensorRT [31] library is used to
optimize inference on its GPU and provides post-training Fig. 4. Sample "hand" images taken by the top camera
quantization for TensorFlow model.
The "empty" class only contains a black background. This class
IV. Methodology tells the bin application that there is nothing in the detection
box. An example is shown in Figure 5.
This section introduces the development of the classification
model on Jetson Nano and K210. We collected 1872 new
recycling waste data to train new models. Experiments were
carried out to evaluate the models’ accuracy and power
consumption.

A. Collect more training data

Firstly, more recycling waste images were collected to increase

the training dataset. As the TrashNet dataset only contains
waste images with white background, making the classification
model vulnerable to variation in lighting conditions and dirt or
contamination in the background. Waste photos with a black Fig. 5. "Empty" images taken by the top camera
background were taken to tell the model the background color
is irrelevant in classification and increase the robustness of
real-world cases. Additionally, we want to add two classes, We collected 1871 images and combined them with the
"empty" and "hand" to the model, but we do not have images TrashNet dataset. Table V lists the number of images in each
belonging to these categories to train the model. Therefore, we class and each dataset.
built a detection model with plywood material and collected a
domain-specific waste dataset.
The waste samples were collected from a nearby recycling bin
and from people directly. Figure 3 shows a waste photo taken
by the camera at the top.
Preprint – A SMART RECYCLING BIN USING WASTE IMAGE CLASSIFICATION AT THE EDGE 6

TABLE V
Number of waste items in each category
Category Number of images in Number of images in
TrashNet our dataset
Cardboard 403 268
Glass 501 130
Paper 594 89
Plastic 482 298
Metal 410 140
Hand 0 920
Empty 0 26

Although recycling waste was divided into five categories, some

of the waste was a mixed waste that contained more than one Fig. 7. MobileNet V1 structure used by K210
group of materials. For example, a coca-cola tin had a plastic
label wrapped around the metal body and a wine glass bottle
had a metal cap. In this case, we labelled the items according to TABLE VI
the largest component of them. Additionally, different countries Training parameters used to build the classification model on Jetson
and recycling companies have different recycling rules. In the Nano
UK, the paperboard materials, such as cartons, are recycled Training Parameter Value
with paper, while in the United States of America (USA), they
belong to cardboard. In this paper, we follow the recycling Learning rate scheduler Constant learning rate scheduler
Train/Validation/Test split ratio 72/18/10
guide of a waste management company in the USA [32]. When Optimizer Adam optimiser
it comes to commercialization in the future, it is crucial to Training epochs 50
consider the recycling rules of the local community. Learning rate 4.3e-05
Fine-tuning training epochs 10
B. Train the models with transfer learning Fine-tuning learning rate 4e-06
Loss functions Sparse categorical cross entropy
The two datasets were used to train new classification models Classifier activation function Softmax
with TensorFlow version 2.9. Firstly, we replaced the Efficient- Include top layers in base model False
Net B0 model layer in the benchmark model with MobileNet Batch size 16
Base model dropout rate 0.2
V3 architecture initialized with the pre-trained weights on
ImageNet. The MobileNet model has a smaller parameter size
and memory usage of the model. Then, the output size of the
C. Build and implement the K210 model
dense layer was changed to seven to add two new classes to
the model. Finally, a resizing layer is inserted between the After building the AI application on Jetson Nano, we used
data augmentation layer and the base model layer to resize the a cheaper and less power-hungry processor, K210, to run
input from 512x384 pixels to 224x224 pixels. By compressing the waste classification model. K210 used its own model
the input after data augmentation, we obtained a clearer image format called kmodel. Sipeed provides sample code to train
for base model input than compressing the data before the classification model based on MobileNet V1 and convert it
data augmentation layer. to kmodel model format with nncase [33]. They applied data
Figure 6 illustrates the structure of our model. The model was augmentation to the image data, as shown in Table VII, and
trained with the hyperparameters shown in Table VI. The final built a model with structure with TensorFlow, shown in Figure
model was converted to the TensorRT model with TensorFlow 7.
version 2.5, and its performance was measured on Jetson Nano.
TABLE VII
Data augmentation used in [33]
Pre-Processing Method Value
Random width and height shift Up to 20%
Random rotation Up to 180◦
Random translation Up to 10%
Random shear Up to 50%

We kept the same data augmentation setup and used the same
training hyperparameters applied in the previous training, as
shown in Table VIII. All training data is resized into 224x224
Fig. 6. Waste classification model structure pixels before training. The model used the same test/validation
split 70/15 as our previous models on Jetson Nano. We first
Preprint – A SMART RECYCLING BIN USING WASTE IMAGE CLASSIFICATION AT THE EDGE 7

V. Results and discussion

The accuracies of models obtained from transfer learning will
be analyzed in this section. Our best model, the 224x224 pixels
MobileNet V3 Large model, achieved an accuracy of 95.98%
in classifying the images into seven groups. The model had
40 IPS on Jetson Nano at 5 W power mode and consumed
30% less power than the EfficientNet B0 model. The K210
model achieved a further 96.64% on waste classification while
the board’s power only drew 0.89 W.

A. Accuracies of models trained with TrashNet dataset on

Jetson Nano
Fig. 8. Photo of the final waste classification device As mentioned in section III-B, The EfficientNet B0 model has
achieved high top-1 accuracy on original test data. However,
we could not use this compute-intensive model directly in the
trained 110 epochs to find the stopping epochs and then trained application. So, we aimed to construct a less resource-intensive
the model with the stopping epochs. model while maintaining higher accuracy than the benchmark
model.
TABLE VIII Firstly, we applied a 70/15/15 training/validation/test split to
Training parameters used to build K210 model
the TrashNet dataset by randomly separating 15% of TrashNet
Training Parameter Value data to form the TrashNet test set. The rest of the TrashNet
Learning rate scheduler Constant learning rate scheduler dataset became a new TrashNet training set. This ratio was used
Optimizer Adam optimiser because the author in [37] demonstrated that both 15% and
Learning rate 4.3e-05 10% could produce the same level of classification accuracy
Loss functions Sparse categorical cross entropy on the TrashNet dataset, which is higher than 20% and 25%
Classifier activation function Softmax with statistical analysis. Other research [38][39] also utilized
Include top layers in base model False 70% of the TrashNet data as training data. So, we consider
Batch size 16 that 70/15/15 would also be suitable for model training.
Dropout rate in base model 0.001
Dropout rate at output 0.001 We trained and tested EfficientNetB0, MobileNet V3 Large
and MobileNet V3 Small with new TrashNet training set using
the hyperparameters in Table VI. All the models converged
We tested both the implementation of MobileNet V1 and V2 on quickly and reached 90% of the highest validation accuracy
K210. K210 only supported MobileNet V1 with alpha equal after 8 epochs. Table IX compares the results we obtained.
to 0.75 and MobileNet V2 with alpha equal to 0.5. OOM As it can be observed, the MobileNet V3 Small model had
error occurs when a larger model is deployed. The ImageNet lower training, validation, and test accuracy than the other
accuracies of the V1 (alpha = 0.75) and V2 (alpha = 0 .5) two models. In contrast, MobileNet V3 large reached similar
were 68.4% [34] and 65.4% [35] correspondingly, so only training accuracy as EfficientNet B0. The test accuracy was
MobileNet V1 architecture was trained. 1.12% higher, while the validation accuracy was 1.38% lower.
The trained model is accelerated into kmodel format and loaded The two models had similar performance on this task while
onto K210 together with the firmware using Kflash provided MobileNet V3 Large was lighter and faster, so MobileNet V3
by Kendryte [36]. We used the firmware that only supports Large was used to build our new model.
basic Application Programming Interface (API) and integrated As the original TrashNet test set only has five classes, 15%
development environment (IDE) on K210 to reduce memory of the "hand" and "empty" group data were separated from
usage. our new dataset and combined with the TrashNet test set to
form the seven-class test set, which evaluated the model’s
D. Measure the performance of the model on the devices performance on the two new class. The rest of our dataset was
added to TrashNet training data and formed the final training
After the classification model was trained, the throughput of data.
each model on Jetson Nano was measured by recording the
latency. Figure 8 demonstrated the final waste detection device. We trained EfficientNetB0 and MobileNet V3 Large with the
The memory usage and power consumption of the device were new training data. The same hyperparameters in Table VI
measured when only the model was running and when the AI were used to train both models. The result is shown in Table
application was running at the same time. The memory usage X. The accuracy on 5-class test set of the EfficientNet B0
was recorded from the system tool of Jetson Nano. We used model increased from 94.40% in Table IX to 95.23%. This
Keweisi voltage current meter to measure the power. It was indicates that the new dataset had a positive effect on the
connected between the power supply and the device to test models and helped them to learn more general features that
the power consumption of Jetson Nano and K210. assist classification tasks. Our photos have different waste
Preprint – A SMART RECYCLING BIN USING WASTE IMAGE CLASSIFICATION AT THE EDGE 8

TABLE IX
Accuracy of three models on 70/15/15 split

Model Total Total keras(.h5) Training Top-1 Validation Top-1 Accuracy

Architecture parameters (M) model size (MB) accuracy (%) Accuracy (%) on test set (%)
EfficientNetB0 4.2 15.8 99.44 96.32 94.40
MobileNet V3 Large 3.0 11.7 99.50 94.94 95.52
MobileNet V3 Small 0.94 3.87 98.94 93.33 93.56

samples, lighting and background, so the accuracy on TrashNet

will not increase significantly.
The data was also fed to MobileNet V3 Large models with base
model resolutions of 512x384 pixels and 224x224 pixels, as we
wanted to reduce the computation resources used. Eventually,
the 22x224 pixels model achieved the highest accuracy on the
TrashNet test set among the three models. The downscaling
increased the accuracy. The 224x224 pixels model maintained
the original test accuracy as the Efficient Net B0 benchmark
model. It achieved a higher per-class precision on both the
original test set and the new test set. This suggested that it was
less biased towards different categories, which was expected
for the application.
The larger input size does not always result in better accuracy.
In [40], the author demonstrated that the misclassification rate
of the model reduced gradually when the resolution reduced
from 2048x2048 pixels to 128x128 pixels. In our case, 224x224
might be a more suitable input size than 512x384, so the Fig. 10. Confusion matrix of MobileNet 224x224
accuracy of MobileNet V3 Large increased when the input
size was reduced. However, more experiments must be carried
out to determine the best resolution for this application. Finally, The model was accelerated into the TensorRT model
and ran on the Jetson Nano. Table XI shows the inference speed
The confusion matrix in Figure 9 and 10 shows that the model of the three models. Compared to the 384x288 EfficientNet
has the highest error rate in classifying between cardboard and model, the MobileNet V3 computation speed increased because
paper. The MobileNet V3 model had the lowest sensitivity of of the reduction of parameter size. The speed doubled for
90% on cardboard against paper, followed by the sensitivity MobileNet V3 with a resolution of 224x224 pixels
of 94.37% on plastic against glass. The EfficientNet and
MobileNet model showed the same characteristics.
TABLE XI
Inference speed on Jetson Nano at maximum power mode
Model Architecture Input Resolution Inference Per
(pixels) second
EfficientNetB0 384x288 19
MobileNet V3 Large 512x384 25
MobileNet V3 Large 224x224 40

To save power, we will run the models at 5 W power mode.

This reduces the computation resources available and limits the
inference speed. Table XII shows that although the inference
speed is reduced to 13 IPS for EfficentNet, the MobileNet
model remains the same inference rate. The memory usage is
reduced to 90.9%.

Fig. 9. Confusion matrix of EffecientNet

Preprint – A SMART RECYCLING BIN USING WASTE IMAGE CLASSIFICATION AT THE EDGE 9

TABLE X
Accuracies of models trained by the final training data

Model Base model Training Top-1 Accuracy Per-class precision Top-1 Accuracy Per-class precision
Architecture input resolution accuracy on 5-class test of 5-class test set on new 7-class of 7-class test set
(pixels) (%) set (%) (%) test set (%) (%)
EfficientNet B0 512x384 99.59 95.23 94.79 95.98 95.97
MobileNet V3 Large 512x384 99.46 94.68 94.53 95.38 95.68
MobileNet V3 Large 224x224 99.46 95.79 95.70 95.98 96.41

TABLE XII
Inference speed on Jetson Nano at 5 W power mode
Model Base model In- Inference Memory Power
Architecture put Resolution Per (%) consumption
(pixels) second (W)
EfficientNetB0 384x288 13 94.93 6.682
MobileNet 224x224 40 90.90 4.698
V3 Large

An AI bin program is programmed using Python Tkinter library.

The program interacts with the user through the graphics user
interface (GUI) and calls the model to perform classifications.
A demonstration video of the program can be found at [41]. The
program occupied 6% of Jetson Nano’s memory, so although it Fig. 11. Change of Model accuracy during training
cannot run EfficientNet B0 model, MobileNet V3 large model
worked fine for it.
At 50 epochs, the training accuracy was still increasing. After
Finally, as shown in Table XII, the power consumption was 100 epochs, it began to oscillate around 97% and stopped
reduced by 29.7% when MobileNet V3 Large model was rising further. So, the training was stopped at 100 epochs.
used. Running the GUI program did not increase the power
consumption. The AI bin program consumed 3.97 W when the The model eventually achieved an accuracy of 96.64% on the
model was not inferencing, and Jetson Nano itself consumed 5-class test set and 95.58% on 7-class test set. The confusion
3 W when no program was running on it. Therefore no matter matrix of the result is shown in Figure 12 . The accuracy of
how light model of the model is, it will still consume 3 W waste classification was higher than the previous models, while
power. Also, it took more than 20 minutes to load the model the 7-class accuracy was 0.5% lower than the MobileNet V3
before inferencing, so Jetson Nano had to keep the program Large 224x224 model. This suggests that K210 could be a
running at 3.97 W. substitute for Jetson Nano on the classification task.
If a 92.5Wh battery pack is used, it will only last 23.3 hours
in standby mode. The battery life will be shorter if the power
consumption of the 10W bulb and the motors are taken into
account. The battery pack must be changed manually every
day if an independent battery source is used. However, the
waste collection frequency in London is once a week.
So, the bin needs to connect to a cable or use a larger and
more expensive power bank instead. This motivates us to test
the performance of K210.

B. Accuracy of the model trained with two datasets on K210

The accuracy of the K210 model was investigated in this

section. The variation in training accuracy during the training
process is shown in Figure 11.

Fig. 12. Confusion matrix of MobileNet V1 on K210

Preprint – A SMART RECYCLING BIN USING WASTE IMAGE CLASSIFICATION AT THE EDGE 10

Fig. 14. Estimated electric energy produced by solar panel in 2019

As it can be observed, the solar panel will have the lowest

Fig. 13. K210 is running the classification model output in November. Assuming that an ideal 48Wh battery is
used to store the power energy produced by the solar panel
and powers the K210 board, the battery can supply 1.9 W
Then, we loaded the model onto the K210 developer kit and power in November ideally, while our application only takes
ran the model. Figure 13 shows the photo of K210 when the 0.89 W. Additionally, it only takes 5 seconds to be switched
model was running on it. The camera connected to it took on and starts detection immediately. This means it will not
photos of the object and the classification result is displayed consume any power in standby mode, significantly reducing
on the top of the screen. The inference time was 66 ms, and the requirements of energy consumption per day. So, using
the device only consumed 0.89W at interference. The power renewable energy to power it may be a feasible choice that
consumption was reduced by 80% compared to the MobileNet can be explored in future studies. The combination of solar
model on Jetson Nano. panels of 1600 cm2 and 48 Wh board would typically cost
between GBP 100 and GBP 200. It would be a cheaper option
We only demonstrated K210’s ability to run the classification than Jetson Nano.
model. In future research, the hyperparameters of the model
can be fine-tuned to improve the model. Additionally, K210
also supports user interface programming with MircoPython, VI. Conclusion and future work
which could be used in the commercialized product.
As the power consumption was less than 1 W, it can be This research aimed to build a real-time waste classification
powered by solar panels in the UK. Photovoltaic Geographical application on Jetson Nano with an accuracy above 95.38%
Information System [42] provided daily and monthly average and a power consumption below 5 W. To eliminate the power
solar radiation data in Europe and other areas through satellites. supply cable connection and further encourage an eco-friendly
The monthly radiation on a plane normal to the sunrays in charging method, K210 was proposed to replace Jetson Nano
London in 2019 was downloaded from the database. The solar for the applications without self-improvement.
panel could be placed at the top of the bin, which will have at The application utilized MobileNet V3 Large as the primary
least 1600 cm2 area. By using a solar panel with an efficiency classification model to reduce the model size and latency.
of 22% [43], the average energy produced a day in an ideal We collected 1800 waste images in the target domain of this
case can be calculated as:[44] application with the detection box. The data was added to
the TrashNet recycling waste dataset to improve the model’s
E = A∗r∗H (1) accuracy. Additionally, two new classes were inserted into
where E is the maximum power output, A equals the solar the classification model that avoids trapping the user’s hand
panel area normal to the sunshine, r is the conversion efficiency, and detects the cases when waste is not present. The model
and H is the monthly irradiation on the plane normal to the eventually achieved a 95.98% accuracy on the test data and
sun rays. Figure 14 shows the electric energy produced by the gave a throughput of 40 IPS at 4.698 W. Finally, we trained
solar panel per day. the MobileNet V1 model with the same training data, and it
reached an accuracy of 96.64%. The model was accelerated
and ran on K210. It only consumed 0.89 W power.
The K210 developer board costs GBP 23, while the exist-
ing intelligent bin product costs USD 5200 [2]. Therefore,
considering the cost of solar batteries, conveyor belts, and
physical shells, the production cost may still be attractive for
commercialization.
Preprint – A SMART RECYCLING BIN USING WASTE IMAGE CLASSIFICATION AT THE EDGE 11

Future works can continue to develop the trash bin with the to improve the recycling performance of urban areas,”
two embedded systems. While Jetson Nano targets the highly Journal of Cleaner Production, vol. 245, p. 118605,
accurate bin used in households and offices, K210 can be used 2020. [Online]. Available: https://www.sciencedirect.com/
in cheaper along-road self-powered bins. For Jetson Nano, science/article/pii/S0959652619334754
the research can investigate the hyperparameters for continual [8] R. Rajkamal, V. Anitha, P. G. Nayaki, K. Ramya, and
training and implement the algorithm on Jetson Nano with E. Kayalvizhi, “A novel approach for waste segregation
limited computation resources that restricts the training time at source level for effective generation of electricity
and samples. For K210, future works can build the circuit board - GREENBIN,” in 2014 International Conference on
for powering K210 with a solar panel. As the energy generated Science Engineering and Management Research, ICSEMR
by the solar panel fluctuates depending on the sunshine, a 2014. Institute of Electrical and Electronics Engineers
power management system is required to ensure constant Inc., February 2014.
voltage output to K210. Moreover, the performance of the
two applications could be tested with a larger unseen real- [9] B. R. S. Kumar, N. Varalakshmi, S. S. Lokeshwari,
world test set with a size greater than 10% of the training data. K. Rohit, Manjunath, and D. N. Sahana, “Eco-friendly
This will give a better estimation of their performance. iot based waste segregation and management,” in 2017
International Conference on Electrical, Electronics, Com-
In conclusion, this paper has provided a state-of-the-art solution munication, Computer, and Optimization Techniques
that reduced the power consumption of Jetson Nano by 30% (ICEECCOT), 2017, pp. 297–299.
and increased the inference speed by 200%. K210 only used
20% of the power of Jetson Nano while maintaining accuracy. [10] W. Pereira, S. Parulekar, S. Phaltankar, and V. Kamble,
The use of this bin can increase the recycling rate up to 95%, “Smart Bin (Waste Segregation and Optimisation),” in
significantly improving the recycling rate in the UK. 2019 Amity International Conference on Artificial Intelli-
gence (AICAI), 2019, pp. 274–279.
[11] D. Ziouzios and M. Dasygenis, “A Smart Recycling Bin
VII. Acknowledgements for Waste Classification,” in 2019 Panhellenic Conference
on Electronics Telecommunications (PACET), 2019, pp.
The authors would like to acknowledge their gratitude to Mindy 1–4.
Yang and Gary Thung who open-sourced the TrashNet dataset.
The authors would also like to show gratitude to the Institute [12] G. White, C. Cabrera, A. Palade, F. Li, and
of Making staff for technical assistance with constructing the S. Clarke, “WasteNet: Waste Classification at the
detection box. Edge for Smart Bins,” June 2020. [Online]. Available:
http://arxiv.org/abs/2006.05873
[13] F. N. Jimeno, B. J. A. Briz, M. R. P. Artiaga, R. E.
References Angelia, and N. B. Limsangan, “Development of Smart
Waste Bin Segregation using Image Processing,” in
[1] S. Kaza, L. C. Yao, P. Bhada-Tata, and F. Van Woerden, 2021 IEEE 13th International Conference on Humanoid,
What a Waste 2.0 : A Global Snapshot of Solid Waste Nanotechnology, Information Technology, Communication
Management to 2050. Urban Development. World and Control, Environment, and Management (HNICEM).
Bank Publications, 2018. [Online]. Available: https: IEEE, November 2021, pp. 1–6. [Online]. Available:
//openknowledge.worldbank.org/handle/10986/30317 https://ieeexplore.ieee.org/document/9732038/
[2] “ Bin-e.” [Online]. Available: https://www.bine.world/ [14] N. C. A. Sallang, M. T. Islam, M. S. Islam, and
[3] Gary Thung and Mindy Yang, “garythung/trashnet: H. Arshad, “A CNN-Based Smart Waste Management
Dataset of images of trash; Torch-based CNN for System Using TensorFlow Lite and LoRa-GPS Shield
garbage image classification.” [Online]. Available: in Internet of Things Environment,” IEEE Access,
https://github.com/garythung/trashnet vol. 9, pp. 153 560–153 574, 2021. [Online]. Available:
https://ieeexplore.ieee.org/document/9615047/
[4] Canaan, “Kendryte K210.” [Online]. Available: https:
//canaan.io/product/kendryteai [15] R. Yamashita, M. Nishio, R. K. G. Do, and K. Togashi,
“Convolutional neural networks: an overview and appli-
[5] Xueying Li, “Paper GitHub Repository.”
cation in radiology,” pp. 611–629, August 2018.
[Online]. Available: https://github.com/molvcan/AN-
SMART-RECYCLING-BIN-USING-WASTE-IMAGE- [16] U. Ozkaya and L. Seyfi, “Fine-tuning models comparisons
CLASSIFICATION-AT-THE-EDGE on garbage classification for recyclability,” CoRR,
vol. abs/1908.04393, 2019. [Online]. Available: http:
[6] “Progress report on recycling and recovery targets
//arxiv.org/abs/1908.04393
for England 2020,” United Kindom Government,
Tech. Rep., January 2022. [Online]. Available: [17] M. Abdulmahmood and R. Grammenos, “Improving the
https://www.gov.uk/government/publications/progress- Deployment of Recycling Classification through Efficient
report-on-recycling-and-recovery-targets-for-england- Hyper-Parameter Analysis,” October 2021. [Online].
2020/progress-report-on-recycling-and-recovery-targets- Available: http://arxiv.org/abs/2110.11043
for-england-2020 [18] Henderi, A. S. Rafika, H. L. H. Spits Warnar, and M. A.
[7] D. Knickmeyer, “Social factors influencing household Saputra, “An Application of Mask Detector for Prevent
waste separation: A literature review on good practices Covid-19 in Public Services Area,” in Journal of Physics:
Preprint – A SMART RECYCLING BIN USING WASTE IMAGE CLASSIFICATION AT THE EDGE 12

Conference Series, vol. 1641, no. 1. IOP Publishing [35] “tensorflow/models/research/slim/nets/mobilenet/.” [On-
Ltd, November 2020. line]. Available: https://github.com/tensorflow/models/
[19] S. Li, K. Huang, L. Feng, Z. Lv, and Y. Wu, tree/master/research/slim/nets/mobilenet
“ADRC-based position control of QCSP with vision [36] Kendryte, “kendryte / kflash.py.” [Online]. Available:
feedback,” in 2021 China Automation Congress (CAC). https://github.com/kendryte/kflash.py
IEEE, October 2021, pp. 5062–5068. [Online]. Available: [37] Y. Yu and R. Grammenos, “Towards artificially
https://ieeexplore.ieee.org/document/9727408/ intelligent recycling: Improving image processing for
[20] A. G. Howard, M. Zhu, B. Chen, D. Kalenichenko, waste classification,” Tech. Rep. [Online]. Available:
W. Wang, T. Weyand, M. Andreetto, and H. Adam, https://arxiv.org/abs/2108.06274
“MobileNets: Efficient Convolutional Neural Networks [38] R. A. Aral, S. R. Keskin, M. Kaya, and M. Haciömeroǧlu,
for Mobile Vision Applications,” April 2017. [Online]. “Classification of TrashNet Dataset Based on Deep Learn-
Available: http://arxiv.org/abs/1704.04861 ing Models,” in Proceedings - 2018 IEEE International
[21] M. Sandler, A. Howard, M. Zhu, A. Zhmoginov, and Conference on Big Data, Big Data 2018. Institute of
L.-C. Chen, “MobileNetV2: Inverted Residuals and Electrical and Electronics Engineers Inc., January 2019,
Linear Bottlenecks,” January 2018. [Online]. Available: pp. 2058–2062.
http://arxiv.org/abs/1801.04381
[39] C. Bircanoglu, M. Atay, F. Beser, O. Genc, and M. A.
[22] A. Howard, M. Sandler, G. Chu, L.-C. Chen, B. Chen, Kizrak, “RecycleNet: Intelligent Waste Sorting Using
M. Tan, W. Wang, Y. Zhu, R. Pang, V. Vasudevan, Q. V. Deep Neural Networks,” in 2018 IEEE (SMC) Interna-
Le, and H. Adam, “Searching for MobileNetV3,” May tional Conference on Innovations in Intelligent Systems
2019. [Online]. Available: http://arxiv.org/abs/1905.02244 and Applications, INISTA 2018. Institute of Electrical
[23] M. Tan and Q. V. Le, “EfficientNet: Rethinking Model and Electronics Engineers Inc., September 2018.
Scaling for Convolutional Neural Networks,” May 2019. [40] M. H. Hesamian, S. Mashohor, M. I. Saripan, and
[Online]. Available: http://arxiv.org/abs/1905.11946 W. A. Wan Adnan, “Effect of image resolution on
[24] N. Ma, X. Zhang, H.-T. Zheng, and J. Sun, “ShuffleNet intensity based scene illumination classification using
V2: Practical Guidelines for Efficient CNN Architecture neural network,” Imaging Science Journal, vol. 63, no. 8,
Design,” Tech. Rep. pp. 433–439, November 2015.
[25] G. Huang, Z. Liu, L. van der Maaten, and [41] Xueying Li, “Demonstration video.” [Online]. Available:
K. Q. Weinberger, “Densely Connected Convolutional https://www.youtube.com/watch?v=_ri6iLEXBhk
Networks,” August 2016. [Online]. Available: [42] European Commission, “PHOTOVOLTAIC GEOGRAPH-
http://arxiv.org/abs/1608.06993 ICAL INFORMATION SYSTEM.” [Online]. Available:
[26] K. He, X. Zhang, S. Ren, and J. Sun, “Deep Residual https://re.jrc.ec.europa.eu/pvg_tools/en/
Learning for Image Recognition,” December 2015. [43] “Lixada 10W Portable Solar Panel High Power
[Online]. Available: http://arxiv.org/abs/1512.03385 Paper Shaped Mini Monocrystalline Silicon Panel
[27] Nvidia, “Jetson Nano Developer Kit.” [Online]. Charger USB Port for Cell Phone Camping
Available: https://developer.nvidia.com/embedded/jetson- Riding Climbing Travel Outdoor Activity.” [Online].
nano-developer-kit Available: https://www.amazon.co.uk/Lixada-Portable-
[28] “Sipeed maixdock datasheet v1.0 key features,” Tech. Monocrystalline-Climbing-Activity/dp/B071D9MB7J/
Rep., 2019. [Online]. Available: www.sipeed.com ref=sr_1_1?geniuslink=true&keywords=lixada+solar+
panel+charger+usb+port+portable+high+power+paper+
[29] TensorFlow, “tensor- shaped+monocrystalline+silicon+for+cell+phone+
flow.org/lite/performance/model_optimization.” [Online]. camping+hiking+travel&qid=1647222985&sr=8-1
Available: https://www.tensorflow.org/lite/performance/
model_optimization [44] Saur News Bureau, “Here is how you can
calculate the annual solar energy output of a
[30] Kendryte, “kendryte / nncase.” [Online]. Available: photovoltaic system,” March 2016. [Online]. Available:
https://github.com/kendryte/nncase https://www.saurenergy.com/solar-energy-blog/here-is-
[31] NVDIA, “NVIDIA TensorRT.” [Online]. Available: how-you-can-calculate-the-annual-solar-energy-output-
https://developer.nvidia.com/tensorrt of-a-photovoltaic-system
[32] L. WM Intellectual Property Holdings,
“Recycling-Myths,” 2020. [Online]. Available:
https://www.wm.com/content/dam/wm/documents/
RecyclingResources/Posters-Guides-Tools/Recycling-
Myths.pdf
[33] Sipeed, “sipeed / maix_train.” [Online]. Available:
https://github.com/sipeed/maix_train
[34] “tensorflow/models/research/slim/nets/mobilenet_v1.md.”
[Online]. Available: https://github.com/tensorflow/models/
blob/master/research/slim/nets/mobilenet_v1.md