A Novel Cross Frequency-Domain Interaction Learning for Aerial Oriented Object Detection

Aerial oriented object detection is a vital task in computer vision, receiving significant attention for its role in remote image understanding. However, most Convolutional Neural Networks (CNNs) methods easily ignore the frequency domain because they only focus on the spatial/channel interaction. To address these limitations, we propose a novel approach called Cross Frequency-domain Interaction Learning (CFIL) for aerial oriented object detection. Our method consists of two modules: the Extraction of Frequency-domain Features (EFF) module and the Interaction of Frequency-domain Features (IFF) module. The EFF module extracts frequency-domain information from the feature maps, enhancing the richness of feature information across different frequencies. The IFF module facilitates efficient interaction and fusion of the frequency-domain feature maps obtained from the EFF module across channels. Finally, these frequency-domain weights are combined with the time-domain feature maps. By enabling full and efficient interaction and fusion of EFF feature weights across channels, the IFF module ensures effective utilization of frequency-domain information. Extensive experiments are conducted on the DOTA V1.0, DOTA V1.5, and HRSC2016 datasets to demonstrate the competitive performance of the proposed CFIL in the aerial oriented object detection. Our code and models will be publicly released.