Size-Adaptive Density Map Estimation and Multiscale Feature Extraction Network for Robust Prawn Postlarvae Counting

Abstract

Accurate prawn postlarvae counting is critical for optimizing aquaculture production and management. Traditional manual counting methods are labor-intensive and prone to errors, while existing automated approaches face challenges, such as small object sizes, overlapping instances, and variations in postlarvae dimensions. To address these limitations, we propose P2CountNet, a novel deep-learning framework that integrates size-adaptive density map estimation with a multiscale feature extraction network for precise and robust prawn postlarvae counting. The size-adaptive density map dynamically adjusts to varying postlarvae sizes, ensuring accurate counting even in dense and heterogeneous environments. Meanwhile, the multiscale feature extraction network captures features across multiple resolutions, effectively detecting tiny and overlapping postlarvae. Experimental evaluations on a newly curated prawn postlarvae dataset demonstrate that P2CountNet significantly outperforms state-of-the-art counting methods, achieving lower mean absolute error and mean squared error. Notably, the proposed method achieves an average counting accuracy of 94.32% for testing datasets. Key contributions include the development of the size-adaptive density estimation technique, the design of a multiscale feature extraction network tailored for small object counting, and extensive evaluations validating the approach. This work provides a robust and efficient solution for prawn postlarvae counting, with promising applications in advancing aquaculture practices.