An Improved Object Recognition Model Based on Deep Learning

Research Abstract

The rapid development of digital platforms and technologies is the reason for spreading a lot of data such as text, videos, sounds, and images. Recently, researchers did a lot of researches in the image processing field such as health, security, social media applications, and so on. But, the recognition of huge amount of multi-classes tiny color images with high accuracy rate becomes one of the most essential problem and need to development. So, this thesis concentrates on large dataset of cifar-10 images which consists of large number of images with different classes. The main task in the image recognition process is the feature extraction phase that distinguishes each image from the other. The high accuracy rate is based on the quality and number of features extracted in the feature extraction phase. Therefore, this thesis is based on developing two models based on combining both traditional and deep learning techniques for feature extraction. In traditional techniques, HOG and SURF are the two used techniques in feature extraction while the used deep learning models are VGG16 and ResNet50. The experimental results show that the combination of HOG, SURF, and ResNet50 is the best combination of features because the accuracy rate reached 98.9% compared with the feature vector of HOG, SURF, and VGG16 which resulted in 98% accuracy rate. Therefore, the accuracy rate that this thesis reached achieved a significant improvement in image recognition field compared with the results of R. Madan et al. [21], S. Srivastava et al. [28], Y. Y. Wang [1], and F. GIUSTE et al. [39] which are 91.09%, 91.1%%, 92% and 94.6% respectively.

Research Keywords

An Improved Object Recognition Model Based on Deep Learning