Swin-TUNA: A Novel PEFT Approach for Accurate Food Image Segmentation

Read full text Summarise with VIRAT
Authors: Haotian Chen, Zhiyong Xiao
Cited by: 0
References: 40

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

In food image processing, parameter-efficient semantic segmentation is important for high-performance applications under constrained training resources. Existing large-scale transformer-based models, such as FoodSAM, still require substantial computational resources during full fine-tuning. This study introduces the TUNable Adapter module (Swin-TUNA), a Parameter-Efficient Fine-Tuning (PEFT) method that integrates multiscale trainable adapters into the Swin Transformer architecture. Swin-TUNA achieves strong food image segmentation performance while updating only 4% of the total parameters. The core innovation is a hierarchical feature adaptation mechanism that uses depth-wise separable convolutions and stage-dependent dimensional mappings to reduce feature mismatch between shallow and deep layers, together with a dynamic balancing strategy for task-agnostic and task-specific features. Experiments show that the proposed method reaches mean Intersection over Union (mIoU) scores of 50.56% and 74.94% on FoodSeg103 and UECFoodPix Complete, respectively, surpassing full fine-tuning of FoodSAM while tuning only 8.13M parameters. In addition, Swin-TUNAdemonstrates faster convergence and stronger discriminative capability, providing an efficient solution for high-precision food image segmentation. The codes are available at https://github.com/chtzs/swin-tuna.

Fields of Study & Subjects

Advanced Neural Network Applications Domain Adaptation and Few-Shot Learning Medical Image Segmentation Techniques

Publication details

Journal / source
IEEE Access
Volume
14
Pages
46118–46128
Publication date
1 January 2026
Language
EN