Fish larvae counting based on feature perception aggregation network

Fish larvae rearing is an essential stage in aquaculture, and accurate fish larvae counting is crucial for fisheries management and ecological conservation. The fish larvae in the rearing stage are often small in size, densely distributed, prone to deformation and occluded between individuals, so the existing counting algorithms have low accuracy for the fish larvae. This results in poor stock management and inefficient resource allocation for aquaculture farms. To address these issues, we propose a Feature Perception Aggregation Network (FPANet) for fish larvae counting. Specifically, we firstly construct the Triple-Attention Refinement Block (TARB) that embeds the Convolutional Block Attention Module (CBAM) to dynamically aggregate contextual information of the images, which not only effectively improves the feature perception of fish larvae with small size but also suppresses shallow noise interference. Secondly, the Dual Cross-attention Feature Enhancement Module (DCFEM) consisting of global branch and local branch is designed. It integrates the cross-attention mechanism to better focus on global information while enhancing local details, effectively addressing fish deformation and occlusion in dense scenarios. Finally, we introduce the dilated convolutions into the decoder of network to enhance the global context information and multi-scale structural information of the images. Experiments show that FPANet achieves an MAE of 1.968, an RMSE of 2.661, and an accuracy of 98.24 %. Compared with the baseline method, FPANet improves accuracy by 0.22 % and reduces MAE and RMSE by 12.8 % and 14.0 %, respectively. Additionally, FPANet outperforms other state-of-the-art methods on our fish larvae dataset. The proposed method can provide support for sustainable aquaculture management and ecological preservation.