Enhancing Aerial Object Detection with Selective Frequency Interaction Network

Aerial object detection is a crucial task in computer vision because it plays a pivotal role in understanding remote images. However, most Convolutional Neural Network (CNN) methods primarily focus on the spatial/channel interactions, overlooking the significance of frequency domain information. To overcome these limitations, we introduce an innovative method named the Selective Frequency Interaction (SFI) network for the task of aerial object detection. Our method comprises two essential modules: the Selective Frequency-domain Feature Extraction (SFFE) module and the Selective Frequency-domain Features Interaction (SFFI) module. In the first module, SFFE, we focus on the extraction of frequency-domain information from the feature maps. This extraction process significantly enriches the feature information, spanning various frequencies. The subsequent module, SFFI, plays a crucial role in facilitating efficient interaction and fusion of the frequency-domain feature maps obtained from the SFFE module across channels. This interaction is essential for optimizing the utilization of frequency-domain information. Finally, we integrate these frequency-domain weights with the time-domain feature maps. By enabling full and efficient interaction and fusion of SFFE feature weights across channels, the SFFI module ensures the effective utilization of frequency-domain information. We conduct extensive experiments on the DOTA V1.0, DOTA V1.5, and HRSC2016 datasets to demonstrate the competitive performance of the proposed SFI network in aerial object detection. The code and model will be available at https://github.com/fzwwj95/EFINet .