Fine-grained food image and text retrieval requires models that can distinguish subtle differences in ingredients, preparation methods, and visual appearance. While vision language models provide strong multimodal reasoning capabilities, their computational requirements can make direct deployment on edge devices impractical. This paper proposes a training free retrieval framework that combines efficient candidate retrieval with selective vision language reasoning. For ingredient retrieval, Qwen3-VL performs closed choice selection over the retrieved candidates, while dense caption retrieval uses Qwen3-VL-Reranker as a second stage multimodal reranker. An efficiency oriented configuration, Stage 4-E, further introduces confidence based execution, allowing the VLM components to be skipped when the initial retrieval is sufficiently confident. The proposed approach is evaluated on the Dishcovery challenge and on auxiliary food retrieval datasets, with particular attention to the tradeoff between retrieval quality and computational cost. The final pipeline is also deployed on two different edge AI platforms, the NVIDIA Jetson AGX Orin and the Qualcomm Dragonwing IQ-9075 EVK, using platform specific model representations and execution backends. The results show that selective multimodal reasoning, confidence based execution, and hardware aware optimization can substantially reduce inference cost while preserving most of the retrieval performance, making the proposed architecture suitable for practical edge deployment.

Spagnolo, A., Bianco, S., Schettini, R., Pau, D. (2026). Edge-Constrained Food Retrieval System. ELECTRONICS, 15(21) [10.3390/electronics15215069].

Edge-Constrained Food Retrieval System

Bianco, Simone;Schettini, Raimondo;
2026

Abstract

Fine-grained food image and text retrieval requires models that can distinguish subtle differences in ingredients, preparation methods, and visual appearance. While vision language models provide strong multimodal reasoning capabilities, their computational requirements can make direct deployment on edge devices impractical. This paper proposes a training free retrieval framework that combines efficient candidate retrieval with selective vision language reasoning. For ingredient retrieval, Qwen3-VL performs closed choice selection over the retrieved candidates, while dense caption retrieval uses Qwen3-VL-Reranker as a second stage multimodal reranker. An efficiency oriented configuration, Stage 4-E, further introduces confidence based execution, allowing the VLM components to be skipped when the initial retrieval is sufficiently confident. The proposed approach is evaluated on the Dishcovery challenge and on auxiliary food retrieval datasets, with particular attention to the tradeoff between retrieval quality and computational cost. The final pipeline is also deployed on two different edge AI platforms, the NVIDIA Jetson AGX Orin and the Qualcomm Dragonwing IQ-9075 EVK, using platform specific model representations and execution backends. The results show that selective multimodal reasoning, confidence based execution, and hardware aware optimization can substantially reduce inference cost while preserving most of the retrieval performance, making the proposed architecture suitable for practical edge deployment.
Articolo in rivista - Articolo scientifico
food image–text retrieval; edge AI; vision–language models; training-free retrieval; SigLIP2; Qwen3-VL; multimodal reranking
English
1-nov-2026
2026
15
21
5069
open
Spagnolo, A., Bianco, S., Schettini, R., Pau, D. (2026). Edge-Constrained Food Retrieval System. ELECTRONICS, 15(21) [10.3390/electronics15215069].
File in questo prodotto:
File Dimensione Formato  
Spagnolo-2026-Electronics-VoR.pdf

accesso aperto

Tipologia di allegato: Publisher’s Version (Version of Record, VoR)
Licenza: Creative Commons
Dimensione 1.98 MB
Formato Adobe PDF
1.98 MB Adobe PDF Visualizza/Apri

I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.

Utilizza questo identificativo per citare o creare un link a questo documento: https://hdl.handle.net/10281/629861
Citazioni
  • Scopus ND
  • ???jsp.display-item.citation.isi??? ND
Social impact