Automatic Parasite Detection on Bivalve Shells Using Detection Transformer

Bivalves, a key blue economy commodity, are highly susceptible to parasitic organisms like worms and barnacles. Farmers, particularly oyster cultivators, face challenges as these parasites reduce product quality and increase mortality rates. To address this, we propose a parasitic worm detection system for bivalves using the Detection Transformer (DETR) with ResNet-50 and ResNet-101 backbones. Our dataset comprises 521 bivalve images 381 from Google Images (secondary data) and 140 primary images categorized into parasite and non-parasite classes. We explored various convolutional dilation strategies to enhance performance. Results show that DETR with ResNet-50 consistently outperforms ResNet-101. The best performance was achieved using dilated convolution at ResNet-50’s stage 4, yielding an AP@50 of 67.7%, AP@75 of 60.0%, mAP of 59.0%, and AR of 61.0%. This approach effectively expanded the receptive field while retaining fine grained features, improving small object detection and contextual understanding. Our findings suggest that ResNet-50, being more efficient, is superior for parasite detection on bivalve shells compared to ResNet-101. Optimal dilation depends on the target object’s visual characteristics, highlighting the need for task specific tuning. This study advances automated parasite detection, enabling faster responses to minimize losses in bivalve aquaculture. Future work should extend the system to detect barnacles and address additional challenges.