Fine-tuned depth-augmented U-Net for enhanced semantic segmentation in indoor autonomous vision systems

Fine-tuned depth-augmented U-Net for enhanced semantic segmentation in indoor autonomous vision systems

연구 분야: Artificial Intelligence

학회: Journal of Real-Time Image Processing

Recent technological advancements have significantly improved indoor autonomous vision systems (IAVSs), underscoring the critical need to enhance their capability to interpret real-world environments in a manner similar to human perception. In response to this challenge, this paper introduces DEADFL-UNet, a groundbreaking framework that enhances the existing EADFL-UNet architecture. EADFL-UNet utilized the EfficientNetB3 model, supplemented by a new Super Attention Block and CBW-FL Loss Function, to tackle the significant data imbalance found in the NYUv2 dataset. Our enhancement focuses on using the MobileNetV2 model in conjunction with several fine-tuning techniques to maximize Depth characteristics in tandem with RGB ones inside the prior architecture. By applying the proposed techniques, we achieved an improvement of approximately 6% in mIoU (Mean Intersection over Union) compared to the original EADFL-UNet model, which was previously published. Furthermore, the difference between the fine-tuned and non-fine-tuned versions is 1.91% in mIoU, demonstrating the significant effectiveness of the fine-tuning technique. To confirm the real-time FPS (Frame Per Second) performance of each model, this technique has undergone extensive testing and assessment using standard metrics, not only on pre-existing datasets but also in a ROS2 (Robot Operating System) simulation environment. These proven techniques have potential for various applications in autonomous systems, such as robotic vision, GPS (Global Positioning System) position tracking, autonomous vehicles, and security, improving accuracy and efficiency.