Detection of Tiny Underwater Feed Remnants and Dynamic Gaussian Density Estimation Based on YOLO–SPD–LSK–DGD

Uneaten feed increases production costs and contributes to water-quality deterioration in shrimp farming, while manual feed-tray inspection is labor-intensive and difficult to sustain for continuous monitoring. Effective feeding assessment requires both residual-feed quantity and spatial distribution, because similar pellet counts can conceal different accumulation patterns. However, existing detection methods struggle with tiny, overlapping pellets and do not directly produce continuous density maps, whereas standalone density estimators are susceptible to underwater background interference. To address these limitations, this study proposes an end-to-end framework that integrates a You Only Look Once (YOLO)-based detector with space-to-depth convolution (SPD-Conv), large selective kernel (LSK) attention, and dynamic Gaussian density (DGD) estimation, termed YOLO–SPD–LSK–DGD. The framework jointly predicts pellet bounding boxes and spatial density maps through a shared backbone. Detail-preserving downsampling and adaptive contextual feature selection strengthen tiny-pellet representations while suppressing interference from tray grids and water-surface reflections. A hybrid bounding-box regression loss combines overlap-based geometric constraints with distribution-based similarity to improve localization stability for densely clustered pellets. Furthermore, density supervision is automatically generated from existing bounding-box centers using bounded Gaussian bandwidths determined by local nearest-neighbor distances. This strategy accommodates variations in pellet spacing without requiring additional manual point annotations. Experiments on a field-collected shrimp feed-tray dataset achieved an mAP50 of 0.9231 for detection and an R2 of 0.9004 for counts obtained by integrating predicted density maps. Evaluation across different accumulation conditions and analysis of failure cases further characterize the model’s performance and limitations under challenging field conditions. These results support the feasibility of jointly locating residual feed and describing its spatial accumulation, providing complementary information for quantitative monitoring and motivating further validation for practical feeding assessment in shrimp aquaculture.