YOLO-Light: Automatic Lightweight You-Only-Look-Once Generation in Different Scenarios Through NeuroEvolution

You-Only-Look-Once (YOLO) represents the state-of-the-art in object detection models. With the emergence of various applications utilizing small domain-specific datasets and limited computing resources for extensive model training and deployment, there is an increasing demand for customized lightweight YOLO architectures. In this paper, we propose a general NeuroEvolution-based method, termed YOLO-Light, designed to automatically create lightweight variants of YOLO architectures tailored to object detection tasks across diverse scenarios. For a given task, YOLO-Light first initializes a population of minimal YOLO architectures and subsequently evolves these models within a novel parallel-chain evolutionary space. This process employs a diversity-protecting evolutionary search strategy until some architectures meet the expected performance standards. During evolution, YOLO-Light incorporates a dynamic evolution regulation mechanism to adjust the evolutionary configuration, thereby enhancing efficiency based on the current evolutionary state. We applied YOLO-Light to generate lightweight YOLOv5, YOLOv8, and YOLOv10 architectures for object detection on the Roboflow 100 small dataset collection, which comprises 100 diverse datasets spanning 7 distinct imagery domains, with a total of 224,714 images and 829 classes. Our experiments focused on 20 datasets ranging from 105 to 8,992 images and 1 to 53 classes. The experimental results show that YOLO-Light reduced the number of parameters by 54–95%, while maintaining or improving mean Average Precision (mAP) compared to standard YOLO architectures. These results demonstrate the effectiveness of YOLO-Light in generating lightweight, task-specific YOLO architectures for resource-constrained object detection tasks. The code repository of YOLO-Light is available on GitHub at https://github.com/BruceShine/YOLO-Light.