DPFA-net: a lightweight hybrid neural network with dual path feature aggregation for food image recognition
文献类型：期刊论文
| 作者 | Zhu, Xiangyi2; Zhang, Wenli2; Sheng, Yingnan2; Lv, Congrui2; Sheng, Guorui2; Min, Weiqing1,3; Jiang, Shuqiang1,3 |
| 刊名 | MULTIMEDIA SYSTEMS
 |
| 出版日期 | 2026-02-03 |
| 卷号 | 32期号:2页码:17 |
| 关键词 | Food image recognition Food computing Lightweight Vision transformer |
| ISSN号 | 0942-4962 |
| DOI | 10.1007/s00530-025-02143-3 |
| 英文摘要 | 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. |
| WOS研究方向 | Computer Science |
| 语种 | 英语 |
| WOS记录号 | WOS:001680937300028 |
| 出版者 | SPRINGER |
| 源URL | [http://119.78.100.204/handle/2XEOYT63/42831]  |
| 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.Chinese Acad Sci, Inst Comp Technol, Key Lab Intelligent Informat Proc, Beijing 100190, Peoples R China 2.Ludong Univ, Sch Comp Sci & Artificial Intelligence, Yantai 264025, Shandong, 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. |
入库方式： OAI收割
来源：计算技术研究所
浏览0
下载0
收藏0
其他版本
除非特别说明，本系统中所有内容都受版权保护，并保留所有权利。