Related Experiment Video
Updated: Mar 9, 2026
Author Spotlight: Assessment of Visual Acuity in Central Vision Loss Through Motion-Based Peripheral Vision Testing
Published on: February 23, 2024
RAC : Few-shot fruit recognition through CLIP-based ambiguity reduction.
Dat Tran-Anh1, Thang Vu Ba2, Ngan Hoang Dao3
1Graduate University of Science and Technology, Vietnam Academy of Science and Technology, Hanoi, Vietnam; Faculty of Information Technology, Thuyloi University, Hanoi, Vietnam.
This study introduces a new framework, Reducing Ambiguity in CLIP for Counterfeit Detection (RAC), to improve the identification of fake agricultural products. RAC effectively reduces visual ambiguities, significantly boosting detection accuracy in few-shot learning scenarios.
Area of Science:
- Computer Vision
- Machine Learning
- Agricultural Science
Background:
- Counterfeit agricultural products present significant classification challenges, especially for models like CLIP under few-shot learning conditions.
- Subtle visual differences between authentic and counterfeit items create ambiguity, hindering accurate detection.
Purpose of the Study:
- To introduce a novel framework, Reducing Ambiguity in CLIP for Counterfeit Detection (RAC), to address ambiguities in counterfeit agricultural product detection.
- To enhance the discriminative power and reduce inter-class confusion in few-shot learning scenarios.
Main Methods:
- The RAC framework utilizes a Multilayer Perceptron Plus (MLP-P) module to synthesize multi-level visual representations from CLIP's image encoder.
- An Inter-class Ambiguity Reduction (IRA) module is employed to suppress visual patterns causing confusion.
- A Network Fusion (NF) stage dynamically integrates outputs for robust classification.
Main Results:
- RAC achieved an average accuracy of 76.95% in 16-shot learning across four benchmarks, outperforming the state-of-the-art by 2.50%.
- The framework reached 80.0% accuracy on the TFS-Fruit dataset, demonstrating its effectiveness in critical agricultural applications.
Conclusions:
- The RAC framework significantly improves the accuracy of counterfeit agricultural product detection, particularly in few-shot learning scenarios.
- RAC's active refinement process effectively mitigates visual ambiguities, offering a robust solution for real-world agricultural challenges.
