Real-time detection and identification of fish skin health in the underwater environment based on improved YOLOv10 model
In densely populated aquaculture net cages, real-time detection and identification of fish skin diseases can effectively prevent large-scale outbreaks, thereby reducing fish mortality rates and economic losses. This study proposes an identification model, DCW-YOLO, based on deep learning-driven obje...
Saved in:
| Main Authors: | , , , , , , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
Elsevier
2025-07-01
|
| Series: | Aquaculture Reports |
| Subjects: | |
| Online Access: | http://www.sciencedirect.com/science/article/pii/S2352513425001097 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | In densely populated aquaculture net cages, real-time detection and identification of fish skin diseases can effectively prevent large-scale outbreaks, thereby reducing fish mortality rates and economic losses. This study proposes an identification model, DCW-YOLO, based on deep learning-driven object recognition technology. By capturing images of fish in a seabed environment, the model can detect fish showing signs of skin diseases. Currently, there is limited research on automatic fish disease recognition specifically for Miichthys miiuy. To address this gap, we introduce, for the first time, a novel dataset for Miichthys miiuy and train, validate, and test the model on this dataset. DCW-YOLO substitutes the CIoU loss function in YOLOv10 with the NWD loss function, thereby improving the model’s ability to detect densely packed targets. The C2f-D-LKA layer is employed in place of the C2f convolutional layer, improving the model’s capacity to capture irregularly shaped and sized objects while effectively reducing computational overhead and parameter load. Additionally, the DySample upsampling structure, which utilizes point sampling, is introduced to increase image resolution without adding significant computational cost. Underwater experimental results show that the mAP50 and precision of DCW-YOLO reach 96.87 % and 95.46 %, respectively, representing improvements of 4.61 % and 3.24 % over the original YOLOv10 model. When deployed in aquaculture settings, this model provides rapid, low-cost real-time disease detection, helping farmers identify diseases early and mitigate potential losses. |
|---|---|
| ISSN: | 2352-5134 |