Related Experiment Video
Updated: Jun 3, 2026
03:57
Anesthesia-free Heartbeat Measurements in Freely Moving Zebrafish
Published on: April 18, 2025
LLM-guided expert feature extraction with fusion-based complementary learning network for fish freshness
Resma Madhu P K1, Rakesh Kumar Sidharthan2, Srinivaas A3
1Department of Electronics and Communication Engineering, Amrita School of Engineering, Amrita Vishwa Vidyapeetham, Coimbatore, India.
Scientific Reports
|June 1, 2026
Summary
This study introduces a novel network integrating expert knowledge and deep learning for accurate fish freshness grading using eye images. The fusion-based complementary learning network achieved 81.03% accuracy, improving non-invasive fish quality assessment.
Area of Science:
- Computer Vision
- Artificial Intelligence
- Food Science
Background:
- Accurate, non-invasive fish freshness grading is challenging, especially in occluded environments.
- Existing deep learning models struggle with inconsistent learning due to lost structural cues in fish eye images.
Purpose of the Study:
- To develop a fusion-based complementary learning network (FCLN) for improved fish freshness classification.
- To integrate expert-suggested photometric features with deep learning visual representations.
Main Methods:
- Utilized a large language model (LLM) to extract numeric features from fish expert knowledge, avoiding manual feature engineering.
- Constructed FCLN using ResNet50 backbone with numeric feature and fusion layers for complementary learning.
- Employed complementary learning analysis (CLA) to validate feature complementarity.
Main Results:
- Achieved a reliable classification accuracy of 81.03% on the Fish Freshness Eye (FFE) dataset.
- CLA demonstrated the complementary nature of numeric features with an average correlation of 0.3755.
- Ablation studies confirmed the importance of combining expert knowledge with deep visual features.
Conclusions:
- The proposed FCLN effectively enhances fish freshness classification by integrating expert knowledge and deep learning.
- The method offers a promising approach for accurate, non-invasive fish quality assessment in challenging conditions.