Journal of Physics: Conference Series
PAPER • OPEN ACCESS
Image classification based on RESNET
To cite this article: Jiazhi Liang 2020 J. Phys.: Conf. Ser. 1634 012110
View the article online for updates and enhancements.
This content was downloaded from IP address [Link] on 24/11/2020 at 15:09
CISAT 2020 IOP Publishing
Journal of Physics: Conference Series 1634 (2020) 012110 doi:10.1088/1742-6596/1634/1/012110
Image classification based on RESNET
Jiazhi Liang
Computer technology, Northwest MuZu University, LanZhou, GanShu, China
*Corresponding author’s e-mail: 601186348@[Link]
Abstract. at present, neural networks are becoming more and more complex, from several layers
to dozens of layers or even more than 100 [Link] main advantage of deep network is that it
can express very complex functions. It can learn features from different levels of abstraction,
such as edge features at lower levels and complex features at higher [Link], the use of
deep networks is not always effective, because there is a very big obstacle - the disappearance of
gradients: in very deep networks, gradient signals tend to approach zero very quickly, which
makes the gradient descent process extremely [Link], in the process of gradient
descent, the weight matrix product operation must be carried out in every step of back
propagation from the last layer to the first layer, so that the gradient will drop exponentially to
0.(in rare cases, there is the problem of gradient explosion, that is, the gradient grows
exponentially to the overflow in the process of propagation). Therefore, in the process of
training, it will be found that with the increase of the number of layers, the rate of gradient
decrease [Link], by deepening the network, although it can express any complex
function, but in fact, with the increase of network layers, we are more and more difficult to train
the network, until the proposal of residual network, which makes it possible to train deeper
network[1].
[Link]
Computer vision recognition is a classic field of artificial intelligence, which has been widely concerned
by academia and industry[2].The cifar-10 data set used in this example is composed of 60000 32 × 32
RGB color images, which are classified into 10 categories, including aircraft, car, bird, cat, elk, dog,
frog, horse, boat and truck, of which 50000 are training pictures and 10000 are test [Link]-100 is
more detailed than [Link] most important feature of cifar-10 data set is that the recognition is
transferred to universal objects and applied to multi [Link]-10 data is stored in a numpy
array of 10 000 × 3072, the unit is uint8s, where 3072 means that a 32 × 32 color image is stored. The
first 1024 bits are R values, the middle 1024 bits are g values, and the last 1024 bits are b values.
[Link]
Deep convolution neural network has a series of major breakthroughs in image classification[3].
However, in the development of deep learning, when we start to consider the convergence of deeper
network, there is a degradation problem, that is, in the deepening neural network, the accuracy will first
rise, and then reach saturation. If the depth is increased, the accuracy will [Link] the error increases
in both training set and test set, it is known that the influence is not caused by over [Link]
enhances its network feature extraction ability through cross layer feature fusion, and network
performance gradually improves with the deepening of [Link] research team tested the deeper
RESNET in an acceptable time, and compared several deep learning models, which proved that
RESNET has better classification performance than other models, and can improve the accuracy by
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
CISAT 2020 IOP Publishing
Journal of Physics: Conference Series 1634 (2020) 012110 doi:10.1088/1742-6596/1634/1/012110
increasing the [Link] (residual neural network) is proposed by he Kaiming team of Microsoft
Research Institute, which is called residual network in [Link] fundamental motivation of
RESNET design is to solve the degradation problem of neural network, that is, when the neural network
is deeper, the training error rate is higher[4]. To solve this problem, the team proposed a residual
[Link] function of network layer is reprogrammed as residual function of input of each [Link]
mathematical statistics, the concept of residual is the difference between the actual observation value
and the estimated value (fitting value).
There are many kinds of RESNET residual components, which can even be defined according to the
project [Link] 2 shows the residual component of resnet-20 used in this paper, which
solves the degradation problem [Link] residual component is composed of two convolution layers and
an identity mapping. The convolution kernel size is 3 ×3. Therefore, the input and output dimensions of
the residual component are the same and can be added [Link] the step size is 1, after batch
regularization, relu activation and convolution of the input in RESNET, the padding layer is the original
input layer; when the step size is 2, the input of RESNET will do the same operation again, and then
average pooling will be carried out to get the filling layer. Finally, the input of output layer is the output
of filling layer plus the output of residual component[5].
Figure 1. residual component of resnet-20.
[Link]-Resnet
In the initiation RESNET module, a 1x1 extended conv operation is added at the end of the initiation
subnet to make its output width (number of channels) the same as the input width of the subnet, so as to
facilitate the addition[6].
[Link]-resnet v1
The revenue RESNET V1 network is mainly used to compare the performance of the inception V3
model. Therefore, the calculation of the inception subnet used by it is reduced compared with the
conventional inception module[7]. This is to ensure that its overall computational / memory overhead is
similar to that of inception v3,Only in this way can we ensure the fairness of the comparison (after all,
Google's view is: the design of CNN deep network is to pursue the performance optimization
under the condition of limited computing and memory)[7]. The pictures below show the inception
RESNET modules used in inception RESNET V1 and the connection modules between them.
2
CISAT 2020 IOP Publishing
Journal of Physics: Conference Series 1634 (2020) 012110 doi:10.1088/1742-6596/1634/1/012110
Figure 2. Inception RESNET_Modules used in V1.
Figure 3. concept RESNET_C module used in V1.
3
CISAT 2020 IOP Publishing
Journal of Physics: Conference Series 1634 (2020) 012110 doi:10.1088/1742-6596/1634/1/012110
Figure 4. concept RESNET_V1 network input module.
[Link] implementation
5.1. Pre requirements
The pandas library, numpy library, opencv and tensorflow (1.0.0 +) need to be pre installed before
running.
5.2. Document organization structure
Among them, cifar10_input.pyIt includes functions of downloading, extracting and preprocessing
cifar-10 [Link] RESNET structure is defined.cifar10_train.pyResponsible for training and
validation.cifar10_test.pyResponsible for testing images..cifar10_main.pyThe starting file for program
execution, including execution training and testing, can start the program by executing this file.
5.3. Parameters
This paper uses imagedatagenerator to enhance data, the importance of data in deep learning is
self-evident, and the contradiction between the lack of data and big data is particularly [Link] this
time, data enhancement is particularly necessary, and the advantages are also well reflected in various
papers;
Figure 5. data enhancement effect.
5.4. Training
Train () defines all classes about the training phase. The main idea is to run train_OPFLAGS.train_Steps
[Link] steps%FLAGS.report_If freq = = 0, it will immediately verify, train, and write all summaries on
the tensorboard.
4
CISAT 2020 IOP Publishing
Journal of Physics: Conference Series 1634 (2020) 012110 doi:10.1088/1742-6596/1634/1/012110
5.5. Testing
The test() function in the train() class will help users predict, and it will return a model
[num_test_images, num_The softmax probability of [Link] user needs to prepare and preprocess the
test data and pass it to the function.
5.6. Experimental results and analysis
5.6.1. Test error
Figure 6. error curve.
5.6.2. The test is accurate
Figure 7. accuracy.
[Link]
As can be seen from the error curve in Figure 7, the training error and test error will be very large in the
early stage of network [Link] the increase of iteration times, the training error and
verification error decreased significantly, from 0.898438 to 0.117188, 0.1280 and 0.348188 in step
9775, from 0.898438 to 0.912 and 2.3543 to 0.117188, 0.1280 and 0.348188 [Link] the
increase of the number of RESNET network layers, the test error of the network in cifar-10 does
decrease. When the number of layers reaches 110, the network performance reaches the [Link] the
RESNET network exceeds a certain level, it is difficult to optimize the network. It may be that the
training network produces over fitting [Link] datasets, using regularization, such as maxout
and dropout, will yield better [Link], in this network, we do not use maxout or dropout, but
simply use regularization to design network architecture.
5
CISAT 2020 IOP Publishing
Journal of Physics: Conference Series 1634 (2020) 012110 doi:10.1088/1742-6596/1634/1/012110
[Link]
This paper proposes a neural network model for image recognition based on deep [Link]
with the traditional vehicle logo recognition method, it integrates multiple features, and can extract
features independently for recognition, thus avoiding the tedious and one-sided feature selection
[Link] experimental results show that this method has good robustness and accuracy, has strong
resistance to noise pollution, and can effectively improve the accuracy of image recognition.
Reasons
The technical terms, paragraph punctuation and quotation format of this paper are accurate and conform
to the academic standards. The full text focuses on the theme and has a clear point of view.
At the same time, this paper maintains a high level of writing while putting forward innovative ideas.
The full text language is concise and accurate, the argument is clear and rigorous, which can better
elaborate and support the views and propositions put forward by him.
Reference
[1] Huang G., Liu S., Maaten L., et al. CondenseNet: An Efficient DenseNet using LearnedGroup
Convolutions. In Conference on Computer Vision and Pattern Recognition, 2752-2761,2018.
[2] Sandler M, Howard A, Zhu M., et al. MobileNetV2: Inverted Residuals and LinearBottlenecks. In
Conference on Computer Vision and Pattern Recognition, 4510-4520, 2018.
[3] Chollet F.. Xception: Deep learning with depthwise separable convolutions. In Conferenceon
Computer Vision and Pattern Recognition, 1800-1807, 2017.
[4] Zhang X., Zhou X., Lin M., et al. Shufflenet: An extremely efficient convolutional neuralnetwork
for mobile devices. In Conference on Computer Vision and Pattern Recognition, 6848-6856,
2018.
[5] Ma N., Zhang X., Zheng H. T., et al. Shufflenet v2: Practical guidelines for efficient
cnnarchitecture design. arXiv preprint arXiv:1807.11164, 2018.
[6] Hu J., Shen L., Sun G.. Squeeze-and-excitation networks. In Conference on Computer Visionand
Pattern Recognition, 7132-7141, 2018.
[7] Zoph B., Vasudevan V., Shlens J., et al. Learning transferable architectures for scalableimage
recognition. In Conference on Computer Vision and Pattern Recognition, 8697-8710,2018.
[8] Liu C., Zoph B., Shlens J., et al. Progressive neural architecture search. arXiv
preprintarXiv:1712.00559, 2017.
[9] G. Wei, H. Ma, W. Qian, et al. Lung nodule classification using local kernel regressionmodels
with out-of-sample extension. Biomedical Signal Processing and Control, 40, 1–9, 2018.