Dual-Path Downsampling Algorithm Based on HWD-MP

In object detection tasks, the quality of downsampling significantly impacts model performance. Conventional downsampling methods, such as max pooling which retains the maximum value within a window, often lead to the loss of critical feature information. This study proposes a plug-and-play dual-path downsampling module, designing the Haar Wavelet-based Downsampling and Max Pooling (HWD-MP) module based on the Haar Wavelet-based Downsampling (HWD) operation. One branch employs max pooling for downsampling, while the another branch utilizes the Haar wavelet transform to decrease the spatial resolution of feature maps while maintaining all the information., thereby addressing the issue of information loss caused by max pooling downsampling and achieving diversity in downsampled features. Experiments on the Yolov7-tiny model demonstrate that this approach can greatly enhance the detection effectiveness, increasing the mAP50 by 2.5% and 1.5% on the RSOD and NWPU VHR-10 remote sensing datasets, respectively.