• Nu S-Au Găsit Rezultate

View of Explore and Analysis of Methods to Train CNN in Machine Learning Environment

N/A
N/A
Protected

Academic year: 2022

Share "View of Explore and Analysis of Methods to Train CNN in Machine Learning Environment"

Copied!
12
0
0

Text complet

(1)

Explore and Analysis of Methods to Train CNN in Machine Learning Environment

ArunaPavate

Dept. of Information Technology Thakur college of Engineering

Mumbai, India

Orcid Id 0000-0002-3387-0905 [email protected]

Dr. Rajesh Bansode Dept. of Information Technology

Thakur college of Engineering Mumbai, India

Orcid Id 000-0002-7054-2601 [email protected]

Dr.A.Prasanth

Dept. of Electronic and CommunicationEngineering,

PSNA College of Engineering and Technology, Dindigul, TamilNadu, [email protected]

Abstract— This work focuses on the use of different course of action to train the deep neural network models. A useful way of thinking about model training from scratch, using transfer learning and most recently using machine learning autoML model is as a variation between simplicity, security, and strength of fit. It is necessary to build the model in the machine learning environment as simple as possible considering the relationship between parameters like time required to train the model, accuracy, reusability of the model, security and size of the dataset. It has been observed that to train the model from the start requires a lot of CPU time, memory, or both and requires more data with respect to the network architecture model, on the other hand transfer learning provides good results as well as computationally efficient but requires knowledge of programming, and data preparation techniques whereas training the model using autoML is cost-effective, intimate knowledge of programming, without extensive model training. Though autoML models have many benefits and have got more attention to train CNN models still need some improvement like data import from a third party, data preparation, feature engineering, etc. This work includes training CNN models from scratch, using transfer learning, and using autoML on the COVID-19 dataset. The models trained and analysed using different factors like time required to train the model, size of the dataset, and accuracy.

Keywords— Transfer learning, COVID-19, CNN, AutoML

I. INTRODUCTION

Nowadays deep learning models have been widely used in many domains from medical image analysis[1][2],education [3], to get the directions for driverless cars[4] and many more. CNN models are having extensive use in many applications designing with revolutionary effect on the real world. Deep learning models like CNN are being used to solve mostly common problems such as image classification in which image samples are classified as per the set of possible classes [5]. To build large and effective classification models, huge amounts of samples(thousands or even millions records) are required to train the model before making a possible prediction. There is no single strategy available to build the model.There are many neural architectural models available from ResNet to CapsNet with enlarging depth and complexity. Training complex models with fewer samples ( less than 5000) leads to the problem of overfitting. To train the model requires both resources and time, which makes building the model more expensive. Training the neural network can be possible using the model build from scratch

(2)

and using transfer learning concepts. Nowadays many producers support building prediction models using a drag and drop facility. This work attempts to catch the performance of the models with building the CNN model from scratch, using fine tuning pretrained models and utilizing the cloud based prediction system to train the model. We can achieve these results on VGG16, ResNet50, DenseNet201 architectures training from scratch, and Imagenet used as baseline systems. For solving classification problems, sample annotation is a very much time consuming task in all domains as samples need to be annotated manually [6]. In practice, it is hard to avail the sufficient amount of data to train the model as well as samples collected in a limited set of conditions , in such cases many researchers uses data augmentation techniques to make minor alterations in available samples such as flipping samples, cropping,translation,rotations of samples [7]. Training the model from scratch focuses on feature extraction and applying data augmentation with the help of that one can prevent the neural network from learning the irrelevant features and increasing the overall performance of the model.

The purpose of this work is to clear with the consequences of different methods used to train the neural network model on safety critical applications specially in medical domain, so current work concentrates on prediction of COVID-19 disease In this work, a model based on Convolutional neural networks is used to detect COVID-19 disease.

Objectives of this work are as follows:

1. To develop a CNN model using training the model from scratch on COIVD -19 dataset

2. To build the model using the transfer learning concept on COIVD-19 dataset to improve learning and model capacity.

3. To develop a model using a cloud based prediction system for COVID-19 dataset (Giotto-AI) .

4. Evaluate the performance of all models to establish the baseline of performance of each model for a classification task.

The remaining part of this work is organized as section 2 discusses the related studies using CNN on COVID-19 X-Ray images for classification considering different training methods. Section 3 describes the proposed methodologies and section 4 evaluates and analyzes experimental results.

II. RELATEDWORK

CNN is the most commonly used machine learning algorithm for learning complex, cognitive classification problems [8][9]. Many researchers have made use of cloud based AI tools for solving the classification, object detection not only in almost all domains. Recently, several researchers have tried an attempt to detect COVID-19 disease using chest x-ray images. Though this disease triggered in December 2019[10] and 29356292 cases confirmed [11] also there is data shortage for COVID-19 chest X-ray images so many researchers used transfer learning to train the model. ChaimaeOuchicha et al.[12]

proposed CVDNet to classify the COIVD-19 disease, this model is trained using a small number of samples and claimed that performance will be improved by training the model with more number of samples. Govardhan et al. [13] developed a system to predict COVID-19 dataset, as the dataset contains only 1215 images data augmentation techniques used to tackle the problem of model overfitting and showed very promising results as 97.77% accuracy.

Wang et al. [14] designed their own dataset using publicly available datasets with 13,975 samples and proposed new architecture COVID-Net. The dataset trained over different architectures VGG- 19,ResNet-50,COVID-Net with claimed highest accuracy 93.3% using COVID-Net and 91.0%

sensitivity. The observed sensitivity is less because of the number of missed COVID-19 samples.

Many other researchers [15][16] trained the model from scratch and claimed it with promising results.

Training the model from scratch and tuning hyperparameters of the network model is a very time consuming task. Building high quality convolutional neural network models which gives the best

(3)

accuracy depends on the human expertise. The solution to this problem as many researchers used pretrainedmodels using the concept of transfer learning.

Loannis et al.[17] proposed a new model CoroNet using the concept transfer learning which basically helps to recognize between COVID-19 Pneumonia and other Pneumonia experimental results showed 89.6% accuracy considering four class cases.

A. Narin et al.[18] used the concept of transfer learning with five different neural network architecture models to classify the COVID-19 disease. The observed results showed the ResNet50 model provided the highest accuracy of 99.7%.

Other related work to train the model using transfer learning to predict the COVID-19 disease through [19] [20][21][22] and many more.

With the use of pretrained models improves the accuracy of the model with less time to train the model but sometimes this ends up with decreasing the performance of the model and is referred to as negative transfer. For knowledge transfer depends on how tasks are related and for that have no specific standards as well as modification of the base model is difficult and changes the number of parameters tuned which is challenging , time consuming.

This study's motivation is to estimate the accuracy of the model after training the models using different methods. This study's main contribution uses a cloud based AI solution (Giotto AI) to forecast the COVID-19 disease and compare the performance of the models to evaluate and find out the effect of training using different methods.

III. METHODOLOGY

This section discussed the various methods used to train the model to predict the COVID-19 disease.

Different modes of training as shown in figure 1:

Fig 1: Modes of training applied on COVID-19 Dataset:

The COVID-19 Xray images dataset[25] applied to predict the disease and models trained using three different training methods. The dataset contains COVID-19 and normal patient image samples having . There are 5853images and two labels (COVID-19/Normal).

1.1. Deep learning Model Building from scratch

Feature extraction using traditional machine learning algorithms have limitations such as computational bottleneck [23], curse of dimensionality [24] as well as requirement of domain and expert knowledge.

Deep neural network is a machine learning architecture with many layers that makes machines learn without being explicitly programmed and presents useful information from learned data. As the available training data day by day deep neural networks become more robust and are applied to solve

Model Training from Scratch Model Training using transfer learning concept Model Training using cloud based prediction tool(Giott o AI)

Model Evaluation to make the system more robust by analyzing various flaws while building the model

(4)

more complicated problems. Figure 2 shows building a deep neural network model for prediction of COVID-19 from the beginning. Convolutional neural network architectures help to extract better features for the classification task at the end. Here in this case a CNN takes a 224,224,3 images and kernel size 3x3 to produce a 2D activation map which measures the similarity between the filter and the training samples. To classify images of COVID-19 following steps were followed:

Fig.2. Model building using Training Input Samples from scratch Basic classification steps include:

1.Import the dataset (COVID-19 dataset)

First, it is required to prepare the training data to provide the network with clean and categorical training samples. These training samples are converted into categorical data(0,1) using the method one-hot encoding. Check the dimension of the dataset i.e. explore the dataset. Need the exact shape of the dataset, check the dimension of the dataset and expand the size of the dataset by creating modified versions of images i.e. data augmentation. Here horizontal flip and rescaling method applied for data augmentation.

2. Build the convolution neural network

To build convolutional neural networks, first define the basic components such as convolutional layer, pooling layer, activation layer, dropout layer, and a Full_connected_layer. The architecture is designed by calling the number of layers. To make our network less complex it is necessary to select the smaller weights than larger. For this we need to adjust the cost function so that while training the magnitude of the weights down, still manage to produce good predictions. In this model dropout rate set to 0.5

3.Select the optimizer

To train the neural network model, Adam optimizer applied. Adam optimizer is different from classical optimizer like stochastic gradient descent which include momentum and learning rate.

4.Use the trained model

Finally train the neural network model 5. Test the model

1.2.Deep learning Model Building using transfer learning

In practice, it is difficult to get the dataset sufficient size to train the model so the model pre-trained on a very large dataset and then the network used for initialization or to extract fixed features of the task of interest. Transfer learning concepts help deep neural networks to train the model faster and require less amount of training time as compared to training from scratch. Training the network concentrates on

(5)

feature extraction whereas in transfer learning shifts towards the complexity of defining the network.

Transfer learning is simply an optimization instead of a training model from scratch ,we use a pretrained model to save time. There is no single strategy in building the model, built model often reused. There are many architectures available with increasing complexity and depth. Training on such models with less images (less than 5000) lead to the problem of overfitting. Here the main idea is that the network trained on different datasets and reused to solve problems in different domains. There are two different approaches used while building the model using transfer learning concept to solve computer vision problems General framework of the transfer learning is represented as shown in equation 1. For COVID- 19 disease prediction objective function is defined as

O= {L,P(L|X)} ={L,η} L={ l1,l2……….,ln}, li ∈ L ---(1)

L is the output space, η is the predictive function that maps corresponding labels for input and output pairs. The objective here is to learn the target conditional probability distribution where input samples are taken from the source domain such that source domain samples and target domain samples are not same. Following steps are applied to build the model.

1. Fine tuning the network model/ Pre-trained Model Approach 1. Select the Source Model

Large and challenging source models are made available by many research institutions. One can choose the pretrained model from available models

2. Reuse the Model:

The available pretrained model can be reused to build the new model as per interest.

3. Tune the Model:

As per task of interest, the model refined on dependent and independent pairs of data and adapted.

Selected sources should be relevant to predictive modeling problems. There must be some relationship between the input and output samples and mapping from input samples to the output samples.

2. Develop Selected Source Model:

From the selected model we can build the more skillful naive model to confirm weight learning performed.

3. Reuse of Model:

The built model fitted on source can be used as the starting point for the next task of interest that is to be designed. Reuse of a model may use the whole model or parts of the model depending upon the methodology used.

4. Model Tuning:

The built model needs to be redesigned on the input-output pair of samples available for the model that is to be built.

To build the present model as shown in figure 3 uses the concept of fine tune the network model as described above. The base model used as Imagenet and four different head models constructed using the network architecture VGG16, ResNet50, DenseNet201, EfficientNetB5.

(6)

Fig. 3. Model building using transfer learning

1.3 Model building using cloud based (Giotto-AI) prediction Model

In this work image classification model is applied to access x-ray images in order to diagnose the COVID-19 disease. Following are the steps applied to build the model as shown in figure 4 :

Fig. 4. Model building using cloud based prediction Model 1. Data selection

In this stage, the type of data is selected to train the model, here we are considering image data( X-ray images). On selected data, two tasks can be possible to perform either segmentation or classification.

The next step,in data selection, is to import the data. It is necessary to verify each class sample and the quality to find out the model performance. The last step in data selection is to visualize a summary of data with details such as size of the dataset, image size etc. All the steps in data selection process are as shown in figure 5 (a)-(c)

.

Fig.5 a) Type of data selection

1. Data Selectio n

2.Preproces sing

3 Model Selectio n

4

Deploymen t

5. Web App

Step 1.1 New Project Step1.2 Data Type Step 1.3 Task Step 1.4 Data Import Step1.5 Visulizatio n

Step 2.1 Data Augmenta tion Step 2.2 Data Visualizati on

Step 3.1 Model selection Step 3.2 Model evaluation Step a.

Metrics Step b Hyperparamet ers selection Step 3.3 Model Visualization

Step 4.1 Docker Step 4.2 WebApp deploym ent

Step 5.1 Test data selection to verify the result Step 5.2 Classify data Step 5.3:

Result analysis reports download s

(7)

Fig.5 b) Type of task selection

Fig.5 c) Data Visualization 2. Preprocessing

Data augmentation is the first step of preprocessing to create artificial images by applying different techniques such as cropping, rotating. Here augmented images are used only to train the model and make sense if it belongs to the same class as the original one. Here samples are increased using augmentation schemes such as horizontal flip, vertical flip, rotate, brightness, random, contrast and zoom as shown in figure 6.

Fig. 6 Data Augmentation 3. Model selection

There is a relationship between model selection, time to perform training and the cost required to predict the result. Selection of models with more number of layers gives better performance. Different models supported by Giotto with transfer learning features as ResNet18, ResNet34, ResNet50,ResNet101,ResNet 152. For this work, the ResNet 34 model selected with a number of epochs 10. Model visualize helps to verify the model selected ,augmentation, performance metrics etc. Model selection and other parameters can be possible to set again depending on the performance of the model.

(8)

4. Deployment

This is the last phase where a trained model is available. Creating docker images which is an optional choice, helps users to integrate solutions with their own solution so that models can be possible to run in any environment. Trained models can be possible to deploy on the web as shown in fig.

Fig. 7. Model deployment : URL:https://cloud.giotto.ai/ic/xcovid19 5. Web App

Trained model interacted using web apps. This phase helps to classify the selected images, visualize and download the predicted results as shown in figure 8.

Fig. 8. Predicted result : Normal IV. Result and Analysis

The proposed work models implemented using google colab and Giotto (Cloud based system). In this work model designed using scratch, using transfer learning concept and using Giotto AI results are verified and analyzed using different parameters like number of parameters, training accuracy, training loss , validation accuracy, validation loss as shown in table 1. The ratio of training samples vs. validation samples considered as 80% and 20% by default Giotto uses the same ratio. Number of epochs are set to 10 for all the models. The confusion matrix generated is shown in figure 10. Giotto COVID-19 evaluation is shown in figure 11 a) Accuracy and 11 b) Loss

Fig.9.Predicted result for one sample model build using training from scratch

(9)

Fig.10. Confusion matrix for the model generated using Giotto

Fig.11. a) Giotto COVID-19 model evaluation using accuracy

Fig.11. b) Giotto COVID-19 model evaluation using Loss

Fig. 12 a) COVID-19 model evaluation using accuracy models trained using transfer learning

(10)

Fig. 12 b) COVID-19 model evaluation using accuracy models trained using transfer learning

Four different models trained using the concept of transfer learning and evaluated using accuracy and loss as shown in figure 12 a) and b) simultaneously.

Table 1: Performance analysis of models trained using different methods

Architectures #para(in million) Train Accuracy Train Loss Val Accuracy Val Loss

Training scratch 57 78 32 87 33

VGG16 14 85 33 87 31

ResNet50 28 80 47 85 42

DenseNet201 68 98 5 96 11

EfficientNetB5 45 61 66 49 71

Giotto (ResNet34) 23 96 12 87 32

Fig. 13 plot of COVID-19 model evaluation using the parameters mentioned in table 1

As per analysis our worst model for validation is EfficientNetB5 with a model accuracy of 49% and our best performance fine tuned model trained used transfer learning showed 96% of accuracy for densenet 201 architecture whereas model build using Giotto-AI trained model analyzed 87% of accuracy. It has been observed that building the model which performs well depends on many factors like number of

(11)

epochs, complexity of the model, augmentation scheme as well as the number of samples used to train the model. Selection augmentation scheme is a difficult task as that should not affect the final predicted result. Though lots of researchers use transfer learning to train the models there are certain questions and answers that need to be explored including base model selection, how it works, transfer bounds and negative transfer. Using AutoML cloud based tools like Giotto it is easy to build the model with less knowledge of coding but still it is necessary to have the knowledge of the domain. Every method has their own strengths and limitations.

V. Conclusion

This work is an attempt in applying different methods to train the neural network models and performance evaluation. Though almost all the models have given good results , it is difficult to design the more robust and trustworthy model, especially considering safety critical applications, as model results depend on many parameters. When working on small dataset, it is beneficial to take advantage of pre-trained model on large dataset to obtain the best results. Model build from scratch showed interesting results but not better compared to transfer learning. This work is entirely exploratory and hopes to offer some insights while constructing safety critical applications. It has been suggested to take guidance from clinical expertise as well as data science expertise while building the model with lots of tools available and help the engineers to build the model without coding. As well as models designed using prediction tools have limitations like source code modification, distribution etc. As data is the main heart of any model, selection of data plays a major role while building the model. In future performance of the models can be evaluated by increasing number of epochs, other evaluation metrics like f1 score, precision, recall, error rate and many more. to conclude the selection of training model.

REFERENCES

[1] Junshui Ma*†, Robert P. Sheridan‡, Andy Liaw†, George E. Dahl§, and Vladimir Svetnik†, " Deep Neural Nets as a Method for Quantitative structure- Activity Relationships", 2015, 55, 2, 263–274 Publication Date:January 30, 2015 https://doi.org/10.1021/ci500747n

[2] Brinker TJ, Hekler A, Utikal JS, Grabe N, Schadendorf D, Klode J, Berking C, Steeb T, Enk AH, von Kalle C," Skin Cancer Classification Using Convolutional Neural Networks: Systematic Review" ,J Med Internet Res 2018;20(10):e11936

[3] Vo C., Nguyen H.P. (2020)," An Enhanced CNN Model on Temporal Educational Data for Program- Level Student Classification", In: Nguyen N., Jearanaitanakij K., Selamat A., Trawiński B., Chittayasothorn S. (eds) Intelligent Information and Database Systems. ACIIDS 2020. Lecture Notes in Computer Science, vol 12033. Springer, Cham. https://doi.org/10.1007/978-3-030-41964-6_38

[4] WaseemRawat,ZenghuiWang,"Deep Convolutional Neural Networks for Image Classification: A Comprehensive Review",Neural Computation 29, 2352–2449 (2017) © 2017 Massachusetts Institute of Technology

[5] S.Jayachitra, A.Prasanth, 'Multi-Feature Analysis for Automated Brain Stroke Classification Using Weighted Gaussian Naïve Baye‘s Classifier', Journal of Circuits, Systems, and Computers, 2021

[6] Jigang Tang, Songbin Li, Peng Liu, "A review of lane detection methods based on deep learning,

Pattern Recognition,Volume 111, 2021,107623,ISSN0031-

3203,https://doi.org/10.1016/j.patcog.2020.107623.

(12)

[7] Iqbal, T., Ali, H. Generative Adversarial Network for Medical Images (MI-GAN). J Med Syst 42, 231 (2018). https://doi.org/10.1007/s10916-018-1072-9

[8] Shorten, C., Khoshgoftaar, T.M. A survey on Image Data Augmentation for Deep Learning. J Big Data 6, 60 (2019). https://doi.org/10.1186/s40537-019-0197-0

[9] J. De Fauw, J. R. Ledsam, B. Romera-Paredes et al., ―Clinically applicable deep learning for diagnosis and referral in retinal disease,‖ Nature Medicine, vol. 24, no. 9, pp. 1342–1350, 2018.

[10] J. De Fauw, J. R. Ledsam, B. Romera-Paredes et al., ―Clinically applicable deep learning for diagnosis and referral in retinal disease,‖ Nature Medicine, vol. 24, no. 9, pp. 1342–1350, 2018

[11] WHO, Pneumonia of unknown cause-China, WHO, 2020,https://www.who.int/csr/don/05-january- 2020-pneumonia-of-unkown-cause-china/en/

[12] V.Balaji, S. Kanaga Suba Raja, S. Aparna, J. Haritha, M. Lakshmi Kanya,‗Smart Assistance for Asperger Syndrome using Raspberry Pi‘, International Journal of Innovative Technology and Exploring Engineering, ISSN: 2278–3075, Volume 9, Issue - 1,2019

[13] Chaimae Ouchicha, Ouafae Ammor, Mohammed Meknassi,CVDNet: A novel deep learning architecture for detection of coronavirus (Covid-19) from chest x-ray images, Chaos, Solitons &

Fractals, Volume 140, 2020,110245,ISSN 0960-0779, https://doi.org/10.1016/j.chaos.2020.110245.

[14] Govardhan Jain and Deepti Mittal and Daksh Thakur and Madhup K. Mittal,"A deep learning approach to detect Covid-19 coronavirus with X-Ray images", Biocybernetics and Biomedical Engineering, volume 40 number 4, pp1391-1405, year 2020, issn 0208-5216, https://doi.org/10.1016/j.bbe.2020.08.008

[15] Wang L., Lin Z.Q. and Wong A. COVID-net: a tailored deep convolutional neural network design for detection of COVID-19 cases from chest x-ray images. arXiv, Mar. 2020.

[16] E. E.-D. Hemdan, M. A. Shouman, and M. E. Karar, ―Covidx-net: A framework of deep learning classifiers to diagnose covid-19 in x-ray images,‖ arXiv preprint arXiv:2003.11055, 2020

[17] J. Zhang, Y. Xie, Y. Li, C. Shen, and Y. Xia, ―Covid-19 screening on chest x-ray images using deep learning based anomaly detection,‖ arXiv preprint arXiv:2003.12338, 2020.

[18] Ioannis D., Apostolopoulos1, T.B. COVID-19: automatic detection from x-ray images utilizing transfer learning with convolutional neural networks. 2020.

[19] A. Narin, C. Kaya, Z. Pamuk, Automatic Detection of Coronavirus Disease (COVID-19) Using X- ray Images and Deep Convolutional Neural Networks

[20] Jamil M, and Hussain I. Automatic Detection of COVID-19 Infection from Chest X-ray using Deep Learning. medRxiv, https://doi.org/10.1101/2020.05.10.20097063, 2020.

[21] Sahinbas K, and Catak FO. Transfer Learning Based Convolutional Neural Network for COVID-19 Detection with X-Ray Images. https://www.ozgurcatak.org/files/papers/covid19- deep-learning.pdf, 2020.

[22] Apostolopoulos ID, Bessiana T. Covid-19: automatic detection from X-ray images utilizing transfer learning with convolutional neural networks. Physical and Engineering Sciences in Medicine 2020;43(635):40. http://dx.doi.org/10.1007/s13246-020-00865-4]

[23] H. S. Maghdid, A. T. Asaad, K. Z. Ghafoor, A. S. Sadiq, and M. K. Khan, ―Diagnosing covid-19 pneumonia from x-ray and ct images using deep learning and transfer learning algorithms,‖ arXiv preprint arXiv:2004.00038, 2020

[24] D. Storcheus, A. Rostamizadeh, and S. Kumar, ―A survey of modern questions and challenges in feature extraction,‖ in Feature Extraction: Modern Questions and Challenges, 2015, pp. 1–18.

[25] Y. Bengio, Y. LeCun et al., ―Scaling learning algorithms towards ai,‖ Large-scale kernel machines, vol. 34, no. 5, pp. 1–41, 2007.

[26] https://www.kaggle.com/bachrr/covid-chest-xray accessed on date sept 2020

[27] Murugan, S., Jeyalaksshmi, S., Mahalakshmi, B., Suseendran, G., Jabeen, T. N., & Manikandan, R.

(2020). Comparison of ACO and PSO algorithm using energy consumption and load balancing in emerging MANET and VANET infrastructure. Journal of Critical Reviews, 7(9), 2020.

Referințe

DOCUMENTE SIMILARE

The proposed work uses the effectiveness of GoogLeNet with minor modifications using transfer learning mechanism in classifying COVID19 patients and normal patients

had suggested machine learning based automatic segmentation and hybrid feature analysis for Diabetic Retinopathy classification using fundus Image (2020).. This study used

In order to develop the financial credit scoring model, various machine learning methods are used.. We propose a machine learning classifier-based analysis model for credit

Thus, after the research work and analyzing the accuracy score and the Coefficient of determination of the different machine learning models, we come to the

This paper has completed in the direction of identifying the spam tweets information using a solitary classifier and hybrid classifier with machine learning

In this work, machine learning methods based on a k- nearest neighbor, support vector machine, naïve Bayes, and random forest classifiers with the integration of genetic algorithm for

Finally, we compare and evaluate few machine learning algorithms in spark using RDD-based regression and classification methods for Random forest, decision tree,

At Present, It Is Very Interesting To Design The Deep Intricate Neural Network (Cnn) Is The Latest Image Recognition Solution. Here We Try To Gather Several