DPFA-net: a lightweight hybrid neural network with dual path feature aggregation for food image recognition

doi:10.1007/s00530-025-02143-3

CSpace

	DPFA-net: a lightweight hybrid neural network with dual path feature aggregation for food image recognition
	Zhu, Xiangyi 1; Zhang, Wenli 1; Sheng, Yingnan 1; Lv, Congrui 1; Sheng, Guorui 1; Min, Weiqing 2,3; Jiang, Shuqiang 2,3
	2026-02-03
发表期刊	MULTIMEDIA SYSTEMS
ISSN	0942-4962
卷号	32 期号:2 页码:17
摘要	Food image recognition holds significant application potential in the field of computer vision. However, due to performance constraints on mobile devices, the scale and computational overhead of models face notable limitations, making effective deployment on mobile platforms challenging. To address this issue, this paper proposes a lightweight Dual-Path Feature Aggregation Network (DPFA-Net), designed to enhance the performance of food recognition tasks through efficient local and global feature extraction strategies. Specifically, the DPFA architecture comprises two core modules: the GhostBottleneck module for local feature encoding and the Position Mamba Vision Transformer (PM-ViT) module for global modeling. In this work, the GhostBottleneck module is utilized to extract local features from images. Furthermore, by integrating the Mamba structure with the Separable Self-Attention (SSA) structure, we construct the Mamba Attention (MA) module, which replaces the traditional Attention mechanism in Vision Transformers to build the PM-ViT module, enabling the capture of global features in food images. The redesigned DPFA-Net effectively fuses local and global information, achieving efficient food image recognition. The experiments were conducted on the ETHZ Food-101, Vireo Food-172, and UEC Food-256 datasets. The results show that, while reducing the number of parameters, DPFA-Net achieved Top-1 accuracies of 91.46%, 91.59%, and 75.33%, respectively, representing a 1.50%-3.9% improvement over MobileViTv2. Compared to MobileViTv2, DPFA-Net improves performance by 1.50-3.9%, fully validating the effectiveness and superiority of the DPFA architecture.
关键词	Food image recognition Food computing Lightweight Vision transformer
DOI	10.1007/s00530-025-02143-3
收录类别	SCI
语种	英语
WOS研究方向	Computer Science
WOS类目	Computer Science, Information Systems ; Computer Science, Theory & Methods
WOS记录号	WOS:001680937300028
出版者	SPRINGER
引用统计
文献类型	期刊论文
条目标识符	http://119.78.100.204/handle/2XEOYT63/42832
专题	中国科学院计算技术研究所
通讯作者	Sheng, Guorui
作者单位	1.Ludong Univ, Sch Comp Sci & Artificial Intelligence, Yantai 264025, Shandong, Peoples R China 2.Chinese Acad Sci, Inst Comp Technol, Key Lab Intelligent Informat Proc, Beijing 100190, Peoples R China 3.Univ Chinese Acad Sci, Beijing 100049, Peoples R China
推荐引用方式 GB/T 7714	Zhu, Xiangyi,Zhang, Wenli,Sheng, Yingnan,et al. DPFA-net: a lightweight hybrid neural network with dual path feature aggregation for food image recognition[J]. MULTIMEDIA SYSTEMS,2026,32(2):17.
APA	Zhu, Xiangyi.,Zhang, Wenli.,Sheng, Yingnan.,Lv, Congrui.,Sheng, Guorui.,...&Jiang, Shuqiang.(2026).DPFA-net: a lightweight hybrid neural network with dual path feature aggregation for food image recognition.MULTIMEDIA SYSTEMS,32(2),17.
MLA	Zhu, Xiangyi,et al."DPFA-net: a lightweight hybrid neural network with dual path feature aggregation for food image recognition".MULTIMEDIA SYSTEMS 32.2(2026):17.