![](bg2.png)
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