Improving the performance of door detection in color images using deep neural networks and data augmentation techniques on various hardware platforms
Phong Bui Hai, Le Minh Hoang, Trần Minh Hoàng, Nguyen Chi Nhon Duc, Phan Dang Huy Khanh, Pham Hoang Lam · 2022 13th International Conference on Information and Communication Technology Convergence (ICTC) · 2022
Door detection is a crucial task of several automatic systems (e.g., robotic navigation, visual impaired support systems). The paper presents the improvements of the door detection in color images. Firstly, various deep neural networks have been investigated to improve the detection accuracy. Secondly, the data augmentation technique is applied to improve the performance of deep neural networks (DNNs). Finally, the door detection is developed on various platforms including the general-purpose graphics processing unit (GPGPU), Jetson nano and Neural Processing Unit (NPU) VIM3 to analyse the computational time and resources. The proposed method has been evaluated and compared with conventional methods on a public dataset (DeepDoor2) to demonstrate the effectiveness and robustness.