计算机科学
软件部署
疾病
自动化
鉴定(生物学)
人工智能
故障排除
钥匙(锁)
医学
特征(语言学)
作者
Iwin John Ettumanukaran,B.V Nagendra,J Joseph Durai Selvam,S. Kumar Chandar
出处
期刊:
[Institute of Electrical and Electronics Engineers]
日期:2026-08-01
卷期号:: 1-6
标识
DOI:10.1109/iconica70605.2026.11736549
摘要
Manual quality inspection of shrimp in seafood export processing suffers from 15-20% inconsistent rejection rates across human inspectors, creating both financial losses from over-rejection and regulatory risk from under-rejection of diseased product. This paper presents the development, evaluation, and real-world deployment of a YOLOv5s-based computer vision system for automated detection of four shrimp diseases-blackgill, White Spot Syndrome Virus (WSSV), pale hepatopancreas, and black spots-at Dolphin Marine Foods & Processors Pvt. Ltd., an EU-approved seafood processing facility in Maharashtra, India. The model was trained on 817 annotated images across the four disease classes using transfer learning from COCO-pretrained weights. Evaluation revealed significant class-dependent performance variation: blackgill detection achieved 80.6% mAP@0.5 with 80% recall, while black spots achieved only 5.1% mAP@0.5 due to visual similarity with natural shell variations. Based on these results, a phased hybrid human-machine deployment strategy was implemented, where the system operates as a screening tool for high-confidence disease classes while human inspectors retain authority over low-confidence cases. The deployed system has inspected over 100 tonnes of product, demonstrating that class-specific deployment strategies can enable practical AI adoption in quality-critical manufacturing even when overall model performance is moderate (44.9% mAP@0.5). This work contributes empirical evidence on deploying imperfect-but-useful computer vision systems in regulated food processing environments.
科研通智能强力驱动
Strongly Powered by AbleSci AI