Automatic Lightweight YOLO Construction through NeuroEvolution for White Shrimps Detection

You Only Look Once (YOLO), as a state-of-the-art computer vision model, is able to realize real-time white shrimp biomass assessment in recirculating aquaculture system if it could be optimized for the computation-limiting environment. In this paper, we propose a NeuroEvolution method to automatically generate lightweight YOLO for real-time white shrimp detection. Our method begins with the initialization of a population of minimal YOLOs from a unique two-level search space, which restricts the overall structure of YOLO and assigns available building blocks. It then employs a constructive evolutionary search strategy to incrementally grow these YOLOs by adding and modifying modules across generations until they meet a preset performance standard. To reduce computational overhead and maintain population diversity during evolution, we design a group scheme that works in tandem with group-centered performance estimation and selection. Our shrimp surveillance experiment results demonstrate that the optimized YOLO models, with fewer than 0.5M parameters, can effectively detect shrimps in real time, achieving approximately 92% mean average precision (mAP) and over 150 frames per second (FPS).