International Conference on Computer and Knowledge Engineering

Home / 14th International Conference on Computer and Knowledge Engineering

Multi-source Ensemble Model for Scene Recognition

Authors :

Amir Hossein Saleknia¹ Ahmad Ayatollahi²

1- School of Electrical Engineering, Iran University of Science and Technology 2- School of Electrical Engineering, Iran University of Science and Technology

Keywords :

Scene Recognition،Ensemble Learning،Test Time Augmentation

Abstract :

In computer vision, scene recognition is one of the most challenging research areas. A possible explanation may be the ambiguity between the various scene classes. Therefore, generating the highly discriminating features necessary for decent performance is a challenging task for a single Deep Neural Network (DNN). In this paper we propose a novel ensemble model for scene recognition. It is composed of multiple DNNs pre-trained for different tasks, each with its own strengths and capabilities. These models complement each other, so using them together can help us achieve higher accuracy than if we used just one alone. In addition, we propose Hard Sample Review (HSR), which is a form of adaptive Test Time Augmentation (TTA) with a lower computation cost and better performance. On the MIT-67 dataset, we achieved 90.15% accuracy by employing these techniques, indicating that the proposed method is superior to previous studies.