Lightweight Semantic Segmentation Model for Disaster Area Based on TransUNet

In this paper, we propose a lightweight and real-time capable variant of the TransUNet model for disaster area semantic segmentation tasks, optimized for deployment on edge devices. While traditional TransUNet models have demonstrated performance in medical image segmentation, their size and complexity hinder real-time inference on resource-constrained platforms. To address this limitation, we used pre-trained R-ViT-Ti_16 model for the encoder part which maintains competitive performance despite a smaller model size. Additionally, we added an extra layer to the decoder and utilized only one skip connection to further reduce complexity. Several experiments were conducted to refine the model, including comparisons of different Vision Transformer (ViT) models, optimizers, loss functions, and activation functions. Our results demonstrated that the combination of R-ViT-Ti_16 with SGD optimizer, Log-Cosh Dice loss, and TanhExp activation function yielded a compact model with 87.5% of the performance of the baseline TransUNet while using only 13% of the model size. The final model was compressed using TensorRT; however, this step introduced significant performance degradation, indicating limitations in achieving real-time efficiency with this particular setup. This research highlights both the potential and the challenges of deploying highly efficient segmentation models on edge devices, suggesting the need for further optimization to balance performance and resource efficiency in real-time applications.