LFCNet: A lightweight fish counting model based on density map regression

Accurate counting of fish is a fundamental task of fish farming and crucial for density regulation, bait planning, and economic benefit evaluation. However, computational capability is a substantial concern for practical aquaculture applications. The paper proposes a lightweight model especially for fish counting applications, termed LFCNet. The model is divided into three components: encoder, decoder, and generation head. Firstly, density map regression is utilized to address the high-density fish issue, making adhesion state counting more accurate. Secondly, Ghost modules are embedded to compress the parameters in mobile device applications. Finally, in order to realize the high-precision counting in a lightweight network, three concentrated-comprehensive convolution (C3) modules and transposed convolution layers are adopted to alleviate the computational overhead and recover the resolution of feature maps, respectively. Experiments show that parameters (16.26 MB) and floating point operations (FLOPs: 69.64G) of LFCNet are reduced by 73.8 % and 64.9 % compared with the baseline method. Meanwhile, the lower root mean square error (RMSE: 6.13) and mean absolute error (MAE: 4.10) indicate that LFCNet has higher counting precision and stability. The generalization studies demonstrate that LFCNet achieves the balance between accuracy and speed in various fish counting scenarios.