YOLO-MMS for aerial object detection model based on hybrid feature extractor and improved multi-scale prediction

Object detection in aerial images has become an important research subject due to the widespread use of aerial platforms, including satellites and unmanned aerial vehicles. However, the task is challenging because it involves a complex background, a high number of small objects, and densely distributed objects, leading to poor detection accuracy. Moreover, despite their excellent detection accuracy, existing one-stage object detection methods have complex structures that require huge computational power, generate high parameters, and exhibit slow inference speed, which makes them unsuitable for edge device applications. To address these issues, this paper proposes an accurate and lightweight object detection model named the YOLO-MMS model. The developed model incorporates several improvements, notably the hybrid backbone structure, which integrates a novel Mix-Mobile inverted bottleneck module to optimize efficiency by reducing the number of generated parameters. Additionally, the multi-scale prediction employs small efficient layer aggregation network and spatial pyramid pooling modules to improve feature extraction across multiple scales. Finally, the model includes an additional detection head and utilizes the Swish activation function to enhance detection accuracy. The evaluation results on the VisDrone and VEDAI datasets demonstrate that the proposed YOLO-MMS model achieved superior accuracy compared to other lightweight YOLO-based models. Furthermore, the proposed model showed significant improvements in model size with a reduction of 41.77% compared to its original YOLOv4-tiny model. These findings indicate that the proposed model presents optimal trade-offs in terms of accuracy and efficiency, rendering it exceptionally suitable for real-time applications on embedded systems. Our code is available at: https://github.com/hanifjunos/YOLO-MMS.