.png){kind=icon}
.png){kind=icon}
Return
High-density fish counting in cages using density-guided cluster optimization network
.png){kind=icon}
PRE
AI
DOI：10.1016/j.eswa.2025.130508.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Fish counting is crucial to balancing stocking density and resource waste in intelligent aquaculture systems. Existing counting methods still suffer from misdetection induced by fish size inconsistency due to fixed receptive fields and insufficient utilization of multi-level features. Herein, this study proposes a point-based fish counting network, Density-Guided Cluster Optimization Network (DGCON), to use cross-resolution interaction and distribution-aware consolidation to achieve precise fish counting in high density cage. In DGCON, multi-resolution interaction leverages the scale complementary fusion to alleviate missed detection of large fish, with Point Distribution Selection Module (PDSM) utilizing sparse density division to discriminate large fish points. Distribution-based point sparsification employs Cluster Refinement Module (CRM) with global proximity merging to mitigate duplicate estimations of small fish. DGCON outperforms the mainstream counting methods in terms of Mean Absolute Error of 7.54 and Mean Square Error of 9.99. Visual heat map proves PDSM effectively highlights high-response sparse regions where large fish exist, and the residual map of CRM demonstrates detection consistency for the same fish through minimal residual fluctuations. DGCON achieves high-precision fish counting in dense cage and enables real-time monitoring of fish population as well as optimization of feeding strategy.
Journal
IF:
7.5
Papers:
2.9W
Citations:
10.2W