Cloud-Aware Fusion of Infrared and RGB Images for Aerial Target Detection

Simultaneously utilizing the complementary information between infrared and RGB images for aerial target detection can improve the detection performance to a certain extent. However, cloud occlusion is difficult to avoid for aerial targets, which limits the upper limit of detection results to some extent. Therefore, in order to solve the problem of cloud occlusion, we have designed corresponding target detection methods for targets in cloud occlusion aerial environments. Specifically, firstly, we propose the Channel and Spatial Dimensions Fusion (CSDF) module, which extracts the differential features of two modalities in parallel from the channel and spatial dimensions, and uses them to enhance its own modality. Secondly, we propose the Convolutional Self Attention (CSA) module and Convolutional Mutual Attention (CMA) module, and use these two modules to form our Cloud-Aware Weight Generation (CAWG) module, which is used to solve the cloud occlusion problem in aerial target detection. We conducted experimental verification on our selfbuilt aerial target dataset, and the experimental results proved the effectiveness of our proposed method.