• A
  • A
  • A
  • АБВ
  • АБВ
  • АБВ
  • А
  • А
  • А
  • А
  • А
Обычная версия сайта

Scientific seminar of the Center for Cognitive Neuroscience "Efficient Adaptation of Vision Foundation Models: Insights from Agricultural Imaging"

12+
Мероприятие завершено

We invite everyone to take part in the scientific seminar of the Center for Cognitive Neuroscience on October 12 at 2:30 PM (Perm time, GMT +5). The speaker is Efrem Yohannes Obsie, Postdoctoral Fellow.

Link to connect online: https://us05web.zoom.us/j/88093131575?pwd=hAFF66PtWKAaiwyDE6uQOZAsZblJ3E.1

Abstract

Vision foundation models provide powerful general-purpose representations for computer vision, but adapting these large pretrained models to specialized domains can require substantial computational resources. This challenge is particularly important in settings where labeled data and computing capacity are limited. This study examines whether vision foundation models can be adapted efficiently to domain-specific image classification without updating all pretrained parameters. We investigate this question in agricultural imaging, where variations in leaf appearance, disease symptoms, and acquisition conditions create a substantial domain shift from the natural images typically used during pretraining.

We propose PLoRA, a parameter-efficient fine-tuning approach that applies Low-Rank Adaptation to the attention layers of pretrained vision transformers. Using the Apple Tree Leaf Disease Segmentation Dataset, we compare linear probing, conventional full fine-tuning, and low-rank adaptation across SwinV2 and DINOv2 models. The best-performing configuration, DINOv2-base with low-rank adaptation, achieves 99.39% classification accuracy while requiring only 1.38 million trainable parameters, compared with 86.78 million under full fine-tuning. This corresponds to a 98.41% reduction in trainable parameters and a 22.47% reduction in training time. The findings show that effective adaptation of large pretrained vision models does not necessarily require full parameter updating. Low-rank adaptation can improve domain-specific performance while substantially reducing computational requirements, highlighting the potential of parameter-efficient fine-tuning for specialized and resource-constrained computer-vision applications.