IMS-YOLOv8: A Lightweight and High-Precision Model for Fine-Grained UAV Detection and Bird Discrimination in Complex Airspace

We propose a lightweight IMS-YOLOv8 model built upon the YOLOv8n architecture for fine-grained detection of multiple drone types and joint discrimination between drones and birds. In the backbone, three novel modules are introduced: MV2Block, a lightweight inverted residual unit that substantially reduces parameters and computational cost; MobileViTBlock, which integrates convolution with local self-attention to enhance global context awareness; and MSCAAttention, a multi-scale convolutional attention module with parallel receptive-field branches that strengthens feature extraction for small targets and complex backgrounds. In the neck, a multi-scale fusion structure is employed, and the detection head makes parallel predictions at three scales P3, P4, P5, outputting bounding boxes and confidence scores for seven drone categories plus birds. At the finest scale, an auxiliary segmentation head produces a binary mask to assist detection. Experiments on a mixed dataset containing seven drone models and birds demonstrate that IMS-YOLOv8 achieves the best performance in Precision, Recall, mAP0.5, and mAP0.5:0.95 compared with YOLOv8, YOLOv9, YOLOv10, and various improved variants, while reducing parameter count to just 1.73 M - 42.6% fewer than the baseline - thereby validating its balance of lightweight design and high accuracy in resource-constrained environments.