A dual-stream ordered topology-aware network for accurate fish counting in aquaculture environments

Accurate fish counting is increasingly important for aquaculture and ecological monitoring, enabling non-invasive population estimates and supporting sustainable management. However, existing methods often fail to maintain accuracy and robustness under challenging conditions such as uneven fish distribution, complex backgrounds, and varying fish morphologies. To address these issues, we propose a Dual-stream Ordered Topology-aware Network (DOT-Net). First, a dual-stream feature embedding framework based on wavelet transform is introduced to encode high- and low-frequency information separately. These features are then aggregated via a dynamic restoration unit, effectively mitigating the long-tail problem caused by uneven fish distribution. Second, an Ordered Filtering Attention mechanism enhances feature representation in fish regions through bidirectional sorting and progressive correlation filtering, suppressing background interference. Third, an Adaptive Topology-aware Strategy is proposed to improve adaptation to morphological variations. It models topological relationships and extracts multilevel features. This enhances structural consistency among individual fish. Experiments show that DOT-Net achieves state-of-the-art performance, with a MAE of 2.48 and RMSE of 3.21 on the Carp Counting Dataset, and 4.07 and 5.50 on the Dense Grass Carp Counting Dataset. The proposed method shows higher accuracy and robustness under varying densities and environments, offering a reliable support for sustainable aquaculture and ecological monitoring.