A method for compressing drone-based small object detection models in complex outdoor scenarios

Abstract

Detecting small targets such as forest fire points in complex outdoor scenes faces challenges including small target size and strong background interference, and is easily affected by motion blur, illumination changes, and complicated backgrounds. Small targets in remote sensing images typically have low pixel coverage and low signal-to-noise ratios, further increasing detection difficulty in complex environments. To address this issue, this paper proposes a model compression method. First, we use the gating weights generated by an adaptive gated feature fusion module as channel importance indicators to perform structured compression on redundant convolutional channels. Second, during the fine-tuning stage after compression, we introduce an NWD loss function. NWD models bounding boxes as 2D Gaussian distributions and uses the Wasserstein distance to measure similarity, which reduces the sensitivity of localization quality evaluation to tiny box overlap changes for small objects. It does not function as a sample weighting mechanism based on IoU. Specifically, when the predicted and ground-truth boxes have no overlap, the gradient of traditional IoU loss becomes zero, while NWD still provides a gradient direction via the Wasserstein distance. Moreover, the NWD distance function is smoother, promoting stable convergence for small object detection., enabling rapid recovery of detection accuracy. Experimental results show that the proposed method mitigates speed degradation caused by increased computational load, achieving a good balance between accuracy and efficiency. While maintaining excellent performance in detecting small targets, our approach significantly improves inference efficiency, making it well-suited for real-time outdoor detection tasks.