the relative position of feature information as an important attribute, while the spatial
change module performs affine transformation on the input image in a dynamic way.
This gives the model a strong deformation, expansion, and rotation invariance.
There is also a network that can improve the recognition ability of an object after
rotation and the recognition between objects. It is a capsule network proposed by
Hinton et al. in 2017 [8]. Traditional CNN is based on neurons, and a neuron represents
an attribute of an object. CapsNet consists of multiple capsules, each consisting of
multiple neurons. Therefore, it has more attribute information than the neural network.
There is an algorithm called dynamic routing between capsules. The main idea is that
the output of the low-level capsules is passed to high-level capsules that can be
modified. If the input of the low-level capsule is very similar to the output of the highlevel capsule, then the two capsules are considered to be highly correlated. The current
capsule network is only three layers shallow, but its performance is better than the
traditional CNN network of the same level. In the future, if we increase the level and
further study the routing algorithm, it will have great development prospects in image
recognition.
In 2019, Zhang et al. proposed a multi-feature weighted network containing two
modules, CFR-DenseNet and ILFR-DenseNet, called MFR-DenseNet [9]. CFRDenseNet adds the SEM module to DenseNet, which is mainly for the channel characteristics of the convolutional layer. ILFR-DenseNet adds DSEM to the front of the
3 Ã 3 convolution layer to simulate the interdependence of different features between
layers. Unlike other CNN models that are superimposed by multiple convolutional
layers, its input is the output of all previous layers. MFR-DenseNet is a network formed
by combining CFR-DenseNet and ILFR-DenseNet. This kind of convergence can
maximize their respective advantages, so that the network can not only explore the
correlation between channels, but also simulate the interdependence of interlayer features. DenseNet is composed of many dense blocks. Its main feature is that each layer
has a relationship with other layers. This ensures maximum information flow, suppresses gradient disappearance during network training, enhances feature propagation,
and is effective reduce model parameters. MFR-DenseNet is the latest research result in
CNN improved model. Due to its special structural features, CNN plays an irreplaceable role in image recognition.
4 Conclusion
Researchers have made significant contributions to deep learning and image recognition. However, at present, deep learning is still not perfect in target detection and image
recognition, and there are still many problems that need to be solved.
The convolutional layer is the most time-consuming when learning on the network.
Studies have shown that the processing time of all convolutional layers accounts for
89% of the total CPU and 95% of the GPU. Therefore, the main research goal of highspeed learning is how to efficient and high-speed performing a large number of multiplication and accumulation calculations. The CPU is mainly used for continuous and
complex operations, while the GPU is better at a large number of parallel operations,
and its application in deep learning will achieve an amazingly high speed. Even so,
520
C. Chen and Y. Tong
change module performs affine transformation on the input image in a dynamic way.
This gives the model a strong deformation, expansion, and rotation invariance.
There is also a network that can improve the recognition ability of an object after
rotation and the recognition between objects. It is a capsule network proposed by
Hinton et al. in 2017 [8]. Traditional CNN is based on neurons, and a neuron represents
an attribute of an object. CapsNet consists of multiple capsules, each consisting of
multiple neurons. Therefore, it has more attribute information than the neural network.
There is an algorithm called dynamic routing between capsules. The main idea is that
the output of the low-level capsules is passed to high-level capsules that can be
modified. If the input of the low-level capsule is very similar to the output of the highlevel capsule, then the two capsules are considered to be highly correlated. The current
capsule network is only three layers shallow, but its performance is better than the
traditional CNN network of the same level. In the future, if we increase the level and
further study the routing algorithm, it will have great development prospects in image
recognition.
In 2019, Zhang et al. proposed a multi-feature weighted network containing two
modules, CFR-DenseNet and ILFR-DenseNet, called MFR-DenseNet [9]. CFRDenseNet adds the SEM module to DenseNet, which is mainly for the channel characteristics of the convolutional layer. ILFR-DenseNet adds DSEM to the front of the
3 Ã 3 convolution layer to simulate the interdependence of different features between
layers. Unlike other CNN models that are superimposed by multiple convolutional
layers, its input is the output of all previous layers. MFR-DenseNet is a network formed
by combining CFR-DenseNet and ILFR-DenseNet. This kind of convergence can
maximize their respective advantages, so that the network can not only explore the
correlation between channels, but also simulate the interdependence of interlayer features. DenseNet is composed of many dense blocks. Its main feature is that each layer
has a relationship with other layers. This ensures maximum information flow, suppresses gradient disappearance during network training, enhances feature propagation,
and is effective reduce model parameters. MFR-DenseNet is the latest research result in
CNN improved model. Due to its special structural features, CNN plays an irreplaceable role in image recognition.
4 Conclusion
Researchers have made significant contributions to deep learning and image recognition. However, at present, deep learning is still not perfect in target detection and image
recognition, and there are still many problems that need to be solved.
The convolutional layer is the most time-consuming when learning on the network.
Studies have shown that the processing time of all convolutional layers accounts for
89% of the total CPU and 95% of the GPU. Therefore, the main research goal of highspeed learning is how to efficient and high-speed performing a large number of multiplication and accumulation calculations. The CPU is mainly used for continuous and
complex operations, while the GPU is better at a large number of parallel operations,
and its application in deep learning will achieve an amazingly high speed. Even so,
520
C. Chen and Y. Tong