# YOLO-Shrimp: A Lightweight Detection Model for Shrimp Feed Residues Fusing Multi-Attention Features

## TL;DR

YOLO-Shrimp is a new lightweight model that detects shrimp feed residues using attention mechanisms and improved loss functions, helping improve efficiency in shrimp farming.

## Contribution

The novel EnSimAM attention mechanism and EnWIoU loss function improve detection of small, dense feed particles in shrimp farming.

## Key findings

- YOLO-Shrimp achieves 70.01% mAP@0.5 and 28.01% mAP@0.5:0.95 on a real-world aquaculture dataset.
- The model reduces parameters by 19.7% and GFLOPs by 14.6% compared to the baseline.
- EnSimAM and EnWIoU enhance detection accuracy for small and densely packed feed particles.

## Abstract

Precise control of feeding rates is critically important in intensive shrimp farming for cost reduction, optimization of farming strategies, and protection of the aquatic environment. However, current assessment of residual feed in feeding trays relies predominantly on manual visual inspection, which is inefficient, highly subjective, and difficult to standardize. The residual feed particles typically exhibit characteristics such as small size, high density, irregular shapes, and mutual occlusion, posing significant challenges for automated visual detection. To address these issues, this study proposes a lightweight detection model named YOLO-Shrimp. To enhance the network’s capability in extracting features from small and dense targets, a novel attention mechanism termed EnSimAM is designed. Building upon the SimAM structure, EnSimAM incorporates local variance and edge response to achieve multi-scale feature perception. Furthermore, to improve localization accuracy for small objects, an enhanced weighted intersection over union loss function, EnWIoU, is introduced. Additionally, the lightweight RepGhost module is adopted as the backbone of the model, significantly reducing both the number of parameters and computational complexity while maintaining detection accuracy. Evaluated on a real-world aquaculture dataset containing 3461 images, YOLO-Shrimp achieves mAP@0.5 and mAP@0.5:0.95 scores of 70.01% and 28.01%, respectively, while reducing the parameter count by 19.7% and GFLOPs by 14.6% compared to the baseline model.