Non-invasive individual identification for precision aquaculture
面向精准水产养殖的非侵入式个体识别
FishFaceID enables accurate, non-invasive individual identification of aquaculture organisms using deep learning.
Our Vim-FFID model achieves 98.81% Acc@1 on sea cucumber identification,
supporting personalized feeding, selective breeding, and health monitoring.
FishFaceID 基于深度学习实现水产养殖生物的精准非侵入式个体识别。Vim-FFID 模型在海参识别上达到 98.81% Acc@1，可支持精准投喂、选择性育种与健康监测。
Abstract / 摘要
In this study, we present FishFaceID, a deep-learning framework for non-invasive individual identification in precision aquaculture (PA). The framework couples a unified evaluation harness with a new identification model, Vim-FFID, tailored to image-based aquaculture organism individual identification. Vim-FFID is a vision mamba-based ID model for fine-grained underwater recognition. It uses class-aware prompt tokens with multi-stage bidirectional prompt–backbone interaction, adds auxiliary supervision on mid-level features, and applies an entropy-gated prototype re-ranking at inference to remain robust under turbidity, occlusion, and cross-session shifts.
We also curate a multi-species, dual-view benchmark covering sea cucumber (Apostichopus japonicus), leopard coral grouper (Plectropomus leopardus), speckled blue grouper (Epinephelus cyanopodus), and grass carp (Ctenopharyngodon idella). On our benchmark, Vim-FFID achieves an accuracy of 98.81% Acc@1 and 99.73% Acc@5 on the Sea Cucumber—Overhead subset.
中文. 本研究提出 FishFaceID：一个面向精准水产养殖（PA）的深度学习非侵入式个体识别框架。该框架将统一的评测流程与新的识别模型 Vim‑FFID 结合，用于基于图像的养殖生物个体识别。Vim‑FFID 是基于 Vision Mamba 的细粒度水下识别模型：通过类别感知的提示 token 与主干网络进行多阶段双向交互，在中层特征上引入辅助监督，并在推理阶段采用熵门控的原型重排序，从而在浑浊、遮挡及跨时段分布变化下保持鲁棒性。我们构建了覆盖海参、豹纹石斑鱼、蓝点石斑鱼与草鱼的多物种双视角基准数据集。在"海参—俯视"子集上，Vim‑FFID 达到 98.81% Acc@1 与 99.73% Acc@5。
Overall Pipeline / 整体流程
FishFaceID integrates three stages: (1) Data Acquisition & Annotation — collecting overhead and underwater images with individual labels; (2) Unified Evaluation Protocol & Model Training — the FishFaceID harness ensures reproducible benchmarking; (3) Vim-FFID Individual Identification — supporting downstream precision aquaculture decision-making such as personalized feeding and selective breeding.
中文. FishFaceID 包含三阶段：① 数据采集与标注——采集俯视与水下图像并标注个体；② 统一评测协议与模型训练——FishFaceID harness 确保可复现的基准评测；③ Vim-FFID 个体识别——支持精准投喂、选择性育种等下游决策。
Vim-FFID Architecture / 模型架构
Vim-FFID is a Vision Mamba-based fine-grained recognition model with four key components:
- CAP (Class-Aware Prompt Bank): Learnable prompt tokens for each class
- CLBI (Prompt-Backbone Cross-Interaction): Multi-stage bidirectional fusion between prompts and backbone features
- IS (Intermediate Supervision): Auxiliary contrastive loss on mid-level features (L=13)
- ER (Entropy-gated Re-Ranking): Prototype-based re-ranking at inference for robust open-set recognition
中文. Vim-FFID 是基于 Vision Mamba 的细粒度识别模型，包含四个核心组件：CAP（类别感知提示库）——为每个类别提供可学习的 prompt token；CLBI（提示-主干交叉交互）——多阶段双向融合；IS（中间监督）——在中层特征上添加辅助对比损失；ER（熵门控重排序）——推理阶段基于原型的重排序，增强开放集鲁棒性。
Dataset Samples / 数据集样例
Our benchmark covers 4 species with dual-view (overhead & underwater) imaging across marine/freshwater habitats.
基准数据集覆盖 4 种物种，包含俯视与水下双视角，涵盖海水/淡水环境。
Sea Cucumber — Overhead / 海参—俯视
Sea Cucumber — Underwater / 海参—水下
Leopard Coral Grouper — Overhead / 豹纹石斑鱼—俯视
Leopard Coral Grouper — Underwater / 豹纹石斑鱼—水下
Speckled Blue Grouper — Overhead / 蓝点石斑鱼—俯视
Speckled Blue Grouper — Underwater / 蓝点石斑鱼—水下
Grass Carp — Overhead / 草鱼—俯视
Grass Carp — Underwater / 草鱼—水下
Results / 结果
Accuracy comparison across 5 species, 3 views (UW/OH/Mixed), and 3 data splits (A1/A2/A3). Vim-FFID (blue line) consistently outperforms all baselines.
对比 5 种物种、3 种视角、3 种数据划分下的准确率。Vim-FFID（蓝线）在各场景下均优于所有基线。
Best Result: Sea Cucumber—Overhead (Five-crop, A3 split): 98.81% Acc@1, 99.73% Acc@5
最佳结果：海参—俯视（Five-crop, A3 划分）：98.81% Acc@1，99.73% Acc@5
BibTeX
FishFaceID
@article{FishFaceID2026,
title = {FishFaceID: A deep learning-based non-invasive system for aquaculture organism individual identification},
author = {Zhang, Qinyue and Shi, Zhensheng and Sun, Naizhe and Wang, Yangfan and Zhang, Lingling and Wang, Bo and Xun, Xiaogang and Zheng, Bing and Zheng, Haiyong},
journal = {Aquaculture},
volume = {613},
pages = {743375},
year = {2026},
doi = {10.1016/j.aquaculture.2025.743375}
}OUC-MOI-ID
@inproceedings{ZhangOUCMOIID2025,
title = {OUC-MOI-ID: A Benchmark Dataset for Marine Organism Individual Identification},
author = {Zhang, Qinyue and Shi, Zhensheng and Sun, Naizhe and Wang, Yangfan and Zhang, Lingling and Zheng, Bing and Zheng, Haiyong},
booktitle = {ICARCAI 2024},
series = {LNNS},
volume = {1376},
pages = {178--189},
year = {2025},
doi = {10.1007/978-981-96-5373-7_15}
}