Page 183 - Kỷ yếu hội thảo khoa học lần thứ 12 - Công nghệ thông tin và Ứng dụng trong các lĩnh vực (CITA 2023)

Tran Quy Nam and Phi Cong Huy 167

To sum up, we simply implement a traditional Convolutional Neural Network with
3 Convolutional layers for feature extraction. The classification is tested with Softmax
function for categorical cross-entropy. Then, we used 4 other classifiers, namely
XGBoost, SVC and Decision Tree Classifier, AdaBoost Classifier, Multi-layer
Perceptron Classifier. The test results of CNN-Softmax, CNN-XGBoost, CNN-SVC,
CNN-Decision Tree Classifier, CNN-AdaBoost, CNN-Multi-layer Perceptron
Classifier models are shown in Table 1 below.

Table 1. Results on accuracy of models

Accuracy (%)
Model
Train Valid Test

CNN-Softmax 71.00 66.00 69.00
CNN-XGBoost 99.98 73.68 72.75

CNN-SVC 75.78 71.18 71.74
CNN-Decision Tree 99.98 66.47 65.94

CNN-AdaBoost 55.58 51.91 52.46
CNN-MLP 87.52 71.03 70.72

We can think that the accuracy is quite low (72.75%). But the research question of
this study is not accuracy, it aims to check whether what kind of hybrid deep learning
has the better performance compare to other hybrid models regarding the problem of
weather image classification. Therefore, we implement a very simple CNN with 3
Convolutional layers for feature extraction which make lower accuracy but run faster.
Our hypothesis does not focus on accuracy but performance of models in
classification. Also, the dataset were imbalance acceptance which leads to lower
accuracy. If we want higher accuracy, we should implement a transfer learning
deeply
trained with million of images in ImageNet database. The aim of this study is to
investigate the hybrid model, which one performs better in term of weather image
classification problem. Therefore, the CNN model to make feature extraction is also
simple; only 3 convolutional layers not include any batch normalization layers as our
purpose is in order to speed up the process of experiments.
The table 1 above shows that the performance of CNN-XGBoost is the best among
other 4 models in the problem of weather image classification. Its values of accuracy
on test set at 72.75% which is higher than other comparative models on the same
tested dataset of weather images.

5 Conclusion

In this study, there were four hybrid models and a simple CNN models, totally five
models, were employed them on the same weather image classification problem. We

ISBN: 978-604-80-8083-9 CITA 2023