A Compact Litopenaeus Vannamei Post-Larvae Fry Counting Network with Optimized Scale Aggregation Network

Abstract

Knowing the number of Litopenaeus vannamei post-larvae fry (PL) is vital in the proliferation process. Traditional methods employ a small spoon for counting which is labor-intensive and has poor accuracy resulting in an uncertain number of biomasses. Slow counting process also contributes to fry hypoxia due to excessive fry contact. This underscores the needs of a compact, high-speed and high-accuracy PL counting network. This paper introduces an Optimized Scale Aggregation Network (OSA-Net), a compact fry counting network based on density map regression, designed for edge devices with a small parameter size of 660 KB. Squeeze-and-Excitation Network embeds the channels pruned network backbone to compress less contributed channels. The model trained with Local Pattern Consistency Loss combined with Euclidean Loss to enhance predicted density map quality. Trained on the Politeknik Elektronika Negeri Surabaya Litopenaeus Vannamei one (PENSLV-1) dataset, proposed model obtained high accuracy with Mean Absolute Error (MAE) of 1.99 and Mean Squared Error (MSE) of 2.69 which indicate superior counting effectiveness in under different density levels and PL sizes.