U-Net + Transformer Encoder and Lora Adaptation for Rural Sheep Segmentation

• Júlio V. M. Marques UFMA / UFPI
• Willians S. Santos UFPI
• Anthony I. M. Luz UFPI
• Rafael L. Araújo UFMA / UFPI
• Alcilene D. de Sousa UFPI
• Jose L. R. Sarmento UFPI
• Romuere R. V. Silva UFMA / UFPI

Resumo

This study presents a lightweight and efficient sheep segmentation method by combining U-Net with Transformerbased encoders and Low-Rank Adaptation (LoRA). RGB videos of 99 sheep were collected using a custom framework with Kinect V2, generating a set of images. Three variants of U-Net were evaluated under identical training conditions. The MiTB5+LoRA model achieved an average IoU of 95.0%±0.04 with a reduced number of trainable parameters. The results show that the proposed approach balances accuracy and efficiency, making it suitable for deployment in resource-limited rural environments.

Palavras-chave: Training, Performance evaluation, Image segmentation, Adaptation models, Accuracy, Computational modeling, Computer architecture, Transformers, Real-time systems, Videos