Page 176 - Kỷ yếu hội thảo khoa học lần thứ 12 - Công nghệ thông tin và Ứng dụng trong các lĩnh vực (CITA 2023)

160

cloudy, rainy, shine, sunrise, snowy, and foggy in dataset. The experiment result with
5-cross validation and 50 epochs showed that the Xception has the best average
accuracy of 90.21% with 10,962 seconds of average training time and MobileNetV2
has the fastest average training time of 2,438 seconds with 83.51% of average
accuracy. Mohamed et al. [3] introduced a novel framework to automatically extract
the information from street-level images relying on deep learning and computer vision
using a unified method without any pre-defined constraints in the processed images.
They designed a pipeline of four deep convolutional neural network models, so-called
WeatherNet, was trained, relying on residual learning using ResNet50 architecture, to
extract various weather and visual conditions such as dawn/dusk, day and night for
time detection, glare for lighting conditions, and clear, rainy, snowy, and foggy for
weather conditions. Their WeatherNet showed strong performance in extracting this
information from user-defined images or video streams. Khan et al. [4] studied some
detection models which were focused on three weather conditions, namely clear, light
snow, and heavy snow, as well as three surface conditions such as dry, snowy,
wet/slushy. They applied them into several pre-trained CNN models, including
AlexNet, GoogLeNet, and ResNet18 with proper modification via transfer learning.
The best performance was achieved using ResNet18 architecture with an
unprecedented overall detection accuracy of 97% for weather detection. Minhas et al.
[5] studied weather prediction from real-world images via targeting classification
using neural networks. In their article, the capabilities of a custom built driver
simulator were explored specifically to simulate a wide range of weather conditions.
The results indicated that the use of synthetic datasets in conjunction with real-world
datasets could increase the training efficiency of the CNNs by as much as 74%.
In other researches on image recognition by CNN combined with other classifier
algorithm, Thongsuwan S. et al [6] proposed a new deep learning model, namely
Convolutional eXtreme Gradient Boosting (ConvXGB) for classification problems
based on convolutional neural network and XGBoost algorithm. They designed their
ConvXGB consists of several convolutional layers to learn the features of the input
and, followed by XGBoost in the last layer for predicting the class labels. In the
testing process, their ConvXGB model was applied into a dataset which were
collected from the University of California at Irvine (UCI) Repository of machine
learning. They concluded that the results of experiments on several datasets showed
that the ConvXGB got slightly better results than CNN and XGBoost alone. In year of
2019, Thiyagarajan, S. [7] investigated the image processing in crack detection in
construction engineering. The author used two-hybrid machine learning models and
classified the concrete digital images. The aim of their research was to classify into
cracks or non-cracks classes of images in concrete digital images. The Convolutional
Neural Network was used to extract features from concrete pictures. And then, they
used these extracted features as inputs for other machine learning models, namely
Support Vector Machines (SVMs) and Extreme Gradient Boosting (XGBoost). The
proposed method was evaluated on a collection of 40,000 real concrete images, and
the experimental results showed that application of XGBoost classifier to CNN
extracted image features included an advantage over SVM approach in the
measurement of the accuracy score. Huang et al. [8] resolved the problem of low

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