Research on E-Commerce Platform Product Image Recognition
Based on ResNet Network
Kunxian Wang
School of Computer Science, China University of Geosciences, Wuhan, China
Keywords: Convolutional Neural Network, Residual Network, E-Commerce
Abstract: With the development and popularity of the Internet and smart mobile devices, online shopping has gradually
become the main way for contemporary people to shop, the massive amount of commodity information makes
it difficult for users to choose, how to get the correct commodity information has become a difficult problem,
and manual labelling of commodity categories is very inefficient. How to improve the correct rate of
commodity classification, numerous academics have conducted extensive research in this field., the current
mainstream method is to use convolutional neural networks for commodity picture classification. In this
paper, we carry out a study on the classification of commodity pictures on e-commerce platforms by
constructing three ResNet networks with different depths, ResNet18, ResNet34 and ResNet50, respectively,
and explore the practical significance by training and comparing the classification accuracy by using the
commodity pictures directly downloaded from the famous e-commerce platforms in China. From the
experimental results, evidently, as the depth of network layers increases, the performance results of the
network are getting better and better, while all three networks have achieved a high accuracy rate, which
indicates that convolutional neural networks have application value in the classification of commodity
pictures.
1 INTRODUCTION
Amidst the swift evolution of the Internet and the
pervasive adoption of electronic devices such as
smartphones and portable computers, individuals'
proclivity to partake in online shopping is discernible.
Over the past few years, a marked surge has been
observed in the diversification of merchandise
accessible for online procurement. As a result, the
challenge of swiftly and accurately locating the
desired products among this vast array of information
has become an urgent matter to address (Wu 2021).
Unlike traditional offline shopping methods, online
shoppers are unable to physically interact with
products. Instead, they rely on the information
provided by sellers, including product images and
accompanying textual details. As e-commerce
continues to grow, the volume of product information
has exploded. Hence, there is a pressing need for
automated product classification to enable users to
efficiently locate their desired items, ultimately
enhancing overall efficiency (Lowe 2004).
Currently, the most widely used method is the
automatic categorization of commodities based on
text keywords.
Merchants need to provide text information to the
commodities when they are on the shelves, and the e-
commerce platform completes the classification of
commodities according to the category information
provided by the merchants. When users enter
keywords to search, the e-commerce platform can
intelligently match the text keywords and recommend
related products. This method is simple to operate and
fast to query, but it requires merchants to provide
accurate commodity information, and this over-
reliance can easily lead to misclassification (Kai-Qi
et al 2014).
Commodity images can provide most of the
commodity information in the most intuitive way, and
if e-commerce platforms can use the classification
method based on commodity images, it can bring a
better user experience to both merchants and users.
Researchers in various countries began to use
traditional machine learning methods for
merchandise image classification many years ago, but
because merchandise images are more complex than
Wang, K.
Research on E-Commerce Platform Product Image Recognition Based on ResNet Network.
DOI: 10.5220/0012804000003885
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 87-91
ISBN: 978-989-758-705-4
Proceedings Copyright © 2024 by SCITEPRESS – Science and Technology Publications, Lda.
87
general dataset images, it is more difficult to extract
features. With the development of deep learning,
researchers started building various network models
for image classification. Convolutional neural
network is a representative network, which has been
widely used in the field of image classification and
achieved better results. In this paper, we will use the
ResNet network in convolutional neural network for
the research of image classification (He et al 2023).
2 METHOD
The experimental method used in this paper is to
obtain image data on the Internet and use
ResNet(Residual Neural Network) network model for
training to obtain a model that can classify
commodity pictures. ResNet has a good performance
in degradation problems and gradient disappearance
problems, which is the main reason for choosing this
model in this paper. Further details will be provided
in the following explanation.
Conventional convolutional neural networks are
confronted with a formidable predicament. As the
quantity of network layers is augmented, the
network's capacity for intricate feature pattern
extraction is enhanced, theoretically leading to
potentially superior outcomes with increased model
depth. Deep networks have been found to suffer from
degradation problems, with network accuracy
saturating or even decreasing when the network depth
increases. The experimental results reveal that the 56-
layer network is not as effective as the 20-layer
network. This phenomenon does not stem from
overfitting, given that the disparity between predicted
and actual values during 56-layer network's training
phase also manifests conspicuously.
Training deep networks poses challenges due to
the problem of degeneracy. In the context of a
shallow network, the introduction of additional layers
in an upward manner to formulate a deep network
may yield an extreme scenario. In this scenario, these
supplementary strata may remain uninformative,
effectively duplicating the shallow network's
features. This implies that the newfound layers
amount to an Identity mapping. In such an instance,
the deep network is expected to exhibit performance
on par with the shallow network, and degradation
should not manifest.
Residual learning was introduced by Kaiming
He's team to address the issue of degradation. In a
stacked layer configuration, denoted as H(x), where x
is the input, the objective is for the model to learn the
residual, as in:
F(x)=H(x)-x. (1)
Hence, the original acquired feature can be
represented as in:
F(x)+x. (2)
The rationale behind this lies in the fact that
learning residuals is simpler compared to directly
learning the original features. When the residual is
eliminated, the stacking layer's function is limited to
a constant mapping. This restriction prevents any
degradation in network performance at a minimum.
Nonetheless, it's crucial to acknowledge that the
residual is rarely, exactly 0. The stacked layers can
acquire novel features from input characteristics due
to the presence of non-zero residuals, which leads to
improved overall performance. Figure 1 illustrates the
structure of residual learning (He et al 2023). This
bears a certain resemblance to a 'short circuit' within
an electrical circuit, thus constituting a shortcut
connection.
Figure 1: Residual learning unit.
In ResNet, one of the residual modules shown in
Figure 1 is called Bottleneck. ResNet is available in
different versions with different numbers of network
layers such as 18, 34, 50, 101, and 152 layers and the
structure of different layers is shown in Table 1 (He et
al 2023).
3 EXPERIMENTAL METHODS
3.1 Experimental Framework
The operating system of this experiment is Windows
10, and the graphics card is GTX-1660ti. there are
many mature open-source frameworks supporting
convolutional neural networks and ResNet residual
networks, such as TensorFlow, PyTorch, Keras,
Caffe, and so on (Abadi et al 2016, Paszke et al 2023
& Jia et al 2014). Pythorch is chosen for this
experiment, and the experimental process and result
analysis are based on the Pytorch framework.
DAML 2023 - International Conference on Data Analysis and Machine Learning
88
Table 1: Resnet Structure of Different Layers.
layer name
output size
18-layer
34-layer
50-layer
101-layer
152-layer
conv1
112×112
7×7, 64, stride 2
conv2_x
56×56
3×3 max pool, stride 2
×2
×3
×3
×3
×3
conv3_x
28×28
×2
×4
×4
×4
×8
conv4_x
14×14
×2
×6
×6
×23
×36
conv5_x
7×7
×2
×3
×3
×3
×3
1×1
average pool, 1000-d fc, softmax
FLOPs
1.8×10
9
3.6×10
9
3.8×10
9
7.6×10
9
11.3×10
9
3.2 Data Description
The data used in this paper comes from Taobao, a
Chinese e-commerce platform, where "Phone",
"Shoes", "Men's clothing" and "Dress" are selected for
product image downloads. There are about 1000
images in each category. The actual number fluctuates
slightly due to the limitation of downloading images
for some products. 80% of all images are used for
training and 20% for testing. Part of the data is shown
in Figure 2.
Since the image size is not the same, use resize to
scale, after the end of processing each image size is
224 * 224.
Figure 2: Partial data presentation.
3.3 Model Construction
The ResNet network part of this experiment was built
by directly calling the library in Pytorch, and
ResNet18, ResNet34, and ResNet50 were used for
comparison experiments. In addition to the network
part, the experimentation also incorporates the
utilization of the Adam optimizer. This optimizer
amalgamates the merits inherent in both the AdaGrad
and RMSProp optimizers. It adeptly facilitates the
adaptive modulation of the learning rate, thereby
expediting the convergence velocity (Kingma and Ba
2017, Duchi et al 2011 & Tieleman and Hinton 2012).
For the loss function, the CrossEntropyLoss function
was chosen, which is a loss function used to calculate
the cross-entropy loss, which takes the labeled data
and the probability distribution of the model output as
inputs and calculates a scalar loss based on the labeled
data and the model's prediction results. The
experimental parameters are presented in Table 2.
Table 2: Experimental Parameters.
Learning rate
Batch_size
Epochs
Device
1e-4
64
30
GPU
4 EXPERIMENTAL RESULTS
AND ANALYSIS
The performance of ResNet18, ResNet34 and
ResNet50 with three different depths of residual
networks on the same dataset is shown in Figure 3,
Research on E-Commerce Platform Product Image Recognition Based on ResNet Network
89
which shows that as the network's layer count
increases, the correctness rate is incrementally
increasing from 85.90% to 88.45%, a more obvious
improvement, which can be reflected in the fact that
the ResNet network successfully solves the problem
of degradation that occurs with the deepening of the
network's layers. This observation further
demonstrates the efficacy of the ResNet architecture
in mitigating the degradation phenomenon that
manifests as the network's layer count increases.
Figure 3: Comparison of ResNet correctness for three
depths.
Figure 4: Partial results.
After outputting the prediction results of the test
set part of the results are shown in Fig. 4. Observation
of the prediction results can be found, for cell phones
and shoes such as a large difference between the goods,
basically, all predicted correctly, but men's clothing
and dresses such as the same clothing category of
goods will be more confusing, it may be that some
skirts are more neutral or some men's clothing is more
ornate. This result indicates that the ResNet network
has practical application for the classification of goods
on the e-commerce platform.
5 CONCLUDING REMARKS
This paper investigates the practical utility of
automated product classification for e-commerce
platforms, where ResNet networks of varying depths
were constructed for experimental purposes, and their
performances were compared. The image dataset used
for experimentation was sourced from prominent e-
commerce shopping websites in China. An important
advantage of the methodology employed in this study
is the elimination of the need for a priori manual
feature extraction, with the models utilized yielding
notably high classification accuracy.
One notable merit of incorporating the ResNet
architecture in this research lies in its inherent
capability to address network degradation and the
issue of gradient vanishing, both common challenges
encountered when training deep neural networks. The
design of ResNet, featuring shortcut connections,
facilitates a smooth flow of gradients during the
backpropagation process, thus rendering the training
process more robust and efficient. This attribute holds
particular significance when dealing with intricate
datasets and deep network architectures.
Nevertheless, despite the evident strengths of
ResNet, some limitations persist in the design of this
article, such as occasional misclassifications of similar
product categories. While this aspect may not be
solely attributed to the architectural choice but may
benefit from further refinement through the fine-
tuning of product categories and the exploration of
advanced techniques for enhancing user-friendly
search optimization. In the foreseeable future, our
research endeavors aim to delve more profoundly into
harnessing the complete potential of ResNet while
conducting further investigations and refinements to
produce more resilient product classification models.
REFERENCES
W. X. TUO Wu, “Fashion Style Recognition Based on
Improved Edge Detection Algorithm”, Journal of
Textile Research, vol. 42, no. 10, pp. 157–162, Oct.
2021.
D. G. Lowe, “Distinctive Image Features from Scale-
Invariant Keypoints,” International Journal of
DAML 2023 - International Conference on Data Analysis and Machine Learning
90
Computer Vision, vol. 60, no. 2, pp. 91–110, Nov.
2004.
H. Kai-Qi, R. Wei-Qiang, T. Tie-Niu, “A Comprehensive
Survey of Image Object Classification and Detection
Algorithms”, Journal of Computer Science and
Technology, vol. 37, no. 6, pp. 1225–1240, 2014.
K. He, X. Zhang, S. Ren, and J. Sun, “Deep Residual
Learning for Image Recognition,” presented at the
Proceedings of the IEEE Conference on Computer
Vision and Pattern Recognition, pp. 770–778.
Accessed: Aug. 23, 2023.
M. Abadi et al., “TensorFlow: Large-Scale Machine
Learning on Heterogeneous Distributed Systems.”
arXiv, Mar. 16, 2016.
A. Paszke et al., “PyTorch: An Imperative Style, High-
Performance Deep Learning Library,” in Advances in
Neural Information Processing Systems, Curran
Associates, Inc., 2019. Accessed: Aug. 24, 2023.
Y. Jia et al., “Caffe: Convolutional Architecture for Fast
Feature Embedding,” in Proceedings of the 22nd ACM
international conference on Multimedia, in MM ’14.
New York, NY, USA: Association for Computing
Machinery, Nov. 2014, pp. 675–678.
D. P. Kingma and J. Ba, “Adam: A Method for Stochastic
Optimization.” arXiv, Jan. 29, 2017.
J. Duchi, E. Hazan, and Y. Singer, “Adaptive subgradient
methods for online learning and stochastic
optimization.,” Journal of machine learning research,
vol. 12, no. 7, pp. 257–269, 2011.
T. Tieleman and G. Hinton, “RMSProp: Divide the
Gradient by a Running Average of its Recent
Magnitude,” University of Toronto, Toronto, 2012.
G. E. Hinton, “Connectionist learning procedures,”
Artificial Intelligence, vol. 40, no. 1–3, pp. 185–234,
Sep. 1989.
Research on E-Commerce Platform Product Image Recognition Based on ResNet Network
91
