Author(s):
1. HAGAR M. El HADAD:
Faculty of Computers and Artificial Intelligence, Department of Information System, Beni-Suef
University, Egypt
2. AHMED ELSAYED YACOUP:
Faculty of Computers and Artificial Intelligence, Department of Information System, Helwan
University, Egypt
3. SHAIMAA S. ABDEL-KADER:
Faculty of Computers and Artificial Intelligence, Department of Information System, Beni-Suef
University, Egypt
4. MANAL A. ABDEL-FATTAH:
Faculty of Computers and Artificial Intelligence, Department of Information System, Helwan University, Egypt
Abstract:
Nowadays, the volume of digital data increasing very rapidly especially the image datasets. The reason behind this increase is the rapid development of digital technologies and platforms such as Facebook and Instagram etc. From this point of view, the researchers started to build applications based on using image classification models. These models used traditional techniques or deep learning techniques to classify the multi-classes images. Most researchers in the field of image classification concluded that there are different problems such as the wrong classification for objects and low accuracy rate value in the case of using many classes that are found as a result of the classification phase. Focusing on the essential problem is recognizing a huge number of images for different classes with a high accuracy rate. This paper presents an improved model in multi-classes images recognition. This model combines traditional techniques with deep learning techniques where the feature vector of these techniques (VGG16+HOG+SURF) or (ResNet50+HOG+SURF) are combined in one feature vector for classification. The fine-tunning method is used to perform classification by the combined feature vectors to classification layers in ResNet50. VGG16 and ResNet50 are examples of deep pre-trained networks while Histogram of Oriented Gradients (HOG) and Speeded Up Robust Features (SURF) are examples of traditional techniques. The experimental results of the presented model in this paper provide an improvement through an excellent accuracy rate when using a combined feature vector of (ResNet50+HOG+SURF) that reached 98.9% for the recognition of the cifar-10 dataset.
Page(s):
308-321
DOI:
DOI not available
Published:
Journal: Journal of Theoretical and Applied Information Technology, Volume: 100, Issue: 2, Year: 2022
Keywords:
deep learning
,
Surf
,
Transfer Learning
,
feature extraction
,
VGG16
,
Cifar10
,
Convolutional Neural Network CNN
,
HOG
,
ResNet50
,
image recognition
References:
References are not available for this document.
Citations
Citations are not available for this document.