ShrimpNet: A Lightweight Model for Efficient Detection and Counting of Shrimp Larvae in Dynamic Aquaculture Environments

Accurate determination of shrimp larvae is crucial as misestimation may lead to overfeeding or underfeeding problems. Traditional methods for aquaculture shrimp larvae counting were inefficient and error-prone due to the limitations associated with manual visual inspections. To tackle these challenges, we introduce ShrimpNet, a novel lightweight and efficient model based on YOLOv8n, designed specifically for detecting tiny shrimp larvae. To enhance feature extraction and effectively fuse local features with channel information, we propose a DWASNet backbone, incorporating depth-wise separable convolution, a channel-wise attention mechanism, and a channel shuffle module, optimizing both detection performance and model efficiency. To further reduce computational complexity while maintaining high detection accuracy, we implemented a dual-path concatenation (DPC) bottleneck and a DPCCSP fusion module that utilizes the cross-stage partial (CSP) network for feature fusion. We also optimized the detection head structure and adjusted feature map sizes, enhancing the model’s ability to detect tiny objects. Moreover, we introduced a novel counting algorithm tailored for different scenarios, ensuring unique and accurate counts in both container-based (still water) and waterslide-based (flowing water) counting systems. Extensive experiments on public and custom shrimp larvae datasets demonstrated that the proposed model outperformed other state-of-the-art models in terms of detection accuracy and parameter quantity.