TFE-NET: A TINY-AWARE FEATURE ENHANCEMENT NETWORK FOR COLLABORATIVE OPTIMIZATION IN SMALL OBJECT DETECTION

In object detection tasks, the sparse distribution, weak saliency and context-dependent nature of small objects pose three major challenges for perception systems. Although end-to-end detection architectures like the YOLO series have achieved a good balance between and speed accuracy in recent years, their utilization of shallow-layer features is low, resulting in performance bottlenecks in micro-object recognition. To address this, this paper proposes a small-object perception-enhanced detection framework, TFE-Net (Tiny-aware Feature Enhancement Network). By constructing a shallow high-resolution feature pathway and a multi-scale fine-grained semantic interaction module, it achieves a lightweight improvement of the YOLOv8s model structure. While maintaining the original model's computational complexity, this method significantly enhances the spatial perception and discrimination accuracy for extremely small objects. Experiments were conducted on the VisDrone dataset. Results show that the improved model boosts the [email protected] from 0.386 to 0.421, with noticeable improvements in PR curves across all categories. This confirms the proposed strategy's ability to perceptually reconstruct and detect small objects in complex scenarios.