Page 177 - 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 161

performance on different datasets on image processing classification, and also aim to
resolve the real situation that there were very few specific large-scale datasets for
training stages on image classification. They proposed a new combination
classification model based on three pre-trained CNN models (VGG19, DenseNet169,
and NASNetLarge) for processing the ImageNet database. The aim was to get better
performance and tried to achieve higher classification accuracy. In their proposed
model, the transfer learning model was based on each pre-trained model which was
constructed as a possible classifier. In the next step, they figured out the best output of
three possible classifiers which was concluded as the final classification choice. The
implementation based on two waste image datasets on the same architectures of the
proposed model had achieved the accuracy score at 96.5% and 94% relatively for
classification problems. It means that their proposed model outperformed several
other methods on image classification problems.

3 Methodology

In this study, we try to implement CNN-XGBoost architecture as the proposed
network to apply for problem of weather image classification. For the combined
model CNN-XGBoost, this study tests the model using the convolutional neural
network for feature extraction and the XGBoost algorithm for image classification
which are applied to the classification of faster images (see Fig. 1).

Fig. 1. XGBoost combined CNN network architecture
At the first stage for the feature extractor, we employ traditional Convolutional Neural
Network model as the primary stage to extract features from the weather images. This
CNN architecture of feature extractor has 3 Convolutional layers, the first with 128
filters, kernel size equal 3 and the other two layers with 64 nodes, each layer is
followed by a Max Pooling with pooling size equal 2 and Dropout layers. The
following layers are a Flatten layer and a Fully Connected layer with 128 nodes. A
convolutional layer acquires a feature series by calculating the production between the
receptive field and kernel. In this network, an activation function, namely Rectified
Linear Unit (ReLU) function is added behind each convolutional layer. The

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