Multiscale Wavelet Pooling Transformer and Dynamic Loss for Small Object Detection

Small object detection plays a crucial role in various applications such as remote sensing, traffic surveillance, and UAV imaging. However, due to the small size, weak features, and frequent occlusion of targets, it remains one of the key challenges in computer vision. To address the limitations of YOLO $\mathbf{v 1 1}$ in terms of limited receptive fields and localization inaccuracies, this paper proposes two improvements: First, a Multiscale Wavelet Pooling Transformer (MWPT) module is introduced to enhance feature extraction and contextual perception. Second, a Scale-aware Dynamic Loss (SD Loss) function is designed to dynamically adjust the weights of localization and scale losses based on object size, improving localization robustness. Experiments conducted on the SOTA-A small object detection dataset show that the improved model achieves gains of $\mathbf{1. 0 \%}$ in [email protected], 0.2% in Precision, and 1.0% in Recall compared to the baseline. The full model consistently outperforms others across all metrics, validating the effectiveness and robustness of the proposed methods in dense and complex scenarios.