Year: 2026
Type: preprint
Abstract: Shrimp disease recognition remains challenging due to small lesion regions, subtle texture changes, similar disease appearances, complex backgrounds, and stringent lightweight deployment requirements. To address these issues, we propose MSLG-Net, a Multi-Scale Local–Global Network built upon EfficientNet-Lite0. Our framework integrates three synergistic modules: a Local Lesion Perception (LLP) module for enhancing subtle lesion signals, a Global Morphology Perception (GMP) module for capturing disease-related spatial morphology, and a Lesion-Guided Local–Global Gated Fusion (LGGF) module for adaptively fusing local cues with global semantic contexts under lesion-aware guidance. Experiments were conducted on two public single-label datasets, BD Fish & Shrimp Disease Dataset and TigerShrimpBD, and a self-constructed ShrimpMTD dual-label dataset. MSLG-Net consistently outperformed the EfficientNet-Lite0 baseline on the two public datasets. On ShrimpMTD, it achieved a Weighted Joint Accuracy of 0.8470, a Species Accuracy of 0.9343, and a Disease Accuracy of 0.8096. Ablation and visualization results further confirm the effectiveness of the proposed modules. Overall, MSLG-Net provides a feasible lightweight solution for shrimp disease recognition in resource-constrained aquaculture scenarios.
Source: Research Square
Institution: Dalian Ocean University
Language: English
Subfield: Water Science and Technology
Field: Environmental Science
Domain: Physical Sciences
SDG: Zero hunger
Open Access status: green
OpenAlex ID: W7236866585
