LE-YOLO: layered enhancement network for tiny object detection

Abstract

Aiming at the challenges in small object detection in aerial images, such as extremely small object scales, weakened features, and background interference, this paper proposes a layered enhancement network named LE-YOLO. The proposed framework initiates enhancement at the feature level by integrating the DySample module, which replaces standard upsampling operators through a dynamic point sampling mechanism. This design significantly improves the reconstruction of fine-grained details with negligible computational overhead. Building upon this improved feature representation, a cross-layer feature fusion module is incorporated to aggregate multi-level features, effectively amplifying semantic coherence in regions of interest while suppressing extraneous background noise. Further refining the detection capability, the model introduces a layered enhancement head architecture. Within this head, a novel semantic information enhancement module employs multi-scale dilated convolutional layers to accommodate the substantial size variation characteristic of small aerial objects. The regression branch is subsequently optimized through receptive field adaptive convolutions, enhancing adaptability to diverse object scales. Complementing these architectural innovations, the training process utilizes the normalized Gaussian Wasserstein distance loss to diminish sensitivity to minor localization inaccuracies, particularly beneficial for small targets. Results demonstrate that compared to YOLOv8s, LE-YOLO improves AP50 by 3.1% and 1.7% on the AITOD and SIMD datasets, and outperforms other advanced models in comparative tests.