Shrimp larvae detection and counting in aquaculture using multiscale feature fusion networks

Accurate detection and counting of shrimp larvae are essential for effective management and optimization of shrimp aquaculture systems. However, existing methods often fail in high-density environments, where larvae exhibit significant size variations and overlap, leading to high error rates and reduced applicability in real-world scenarios. To address these challenges, we propose a novel framework that integrates size-adaptive density map generation with a multiscale feature fusion network (MFFN), specifically designed for the dense and complex conditions of aquaculture. The proposed approach dynamically adjusts density map generation based on the varying sizes of shrimp larvae within an image, improving precision and enhancing detection accuracy. Simultaneously, the MFFN architecture extracts critical features across multiple scales, enabling the model to accurately detect and count larvae even in crowded scenes with significant occlusion. By combining these techniques, the framework demonstrates superior adaptability and robustness across diverse aquaculture environments. Extensive evaluations on a comprehensive shrimp larvae dataset show that our method achieves a detection accuracy of 93.68%, significantly outperforming traditional approaches. The model also reduces error rates and improves precision and recall metrics, with visual results showcasing its ability to distinguish and accurately count larvae of varying sizes. This work provides a valuable solution for automated shrimp larvae monitoring, supporting sustainable aquaculture practices. Future research will explore broader applications of the model and further refinements for enhanced scalability.