Learning to Reduce Information Bottleneck for Object Detection in Aerial Images

Object detection in aerial images is a critical and essential task the fields of geoscience remote sensing. Despite popularity computer vision methods detecting objects, these have been faced with significant limitations such as appearance occlusion variable object sizes. In this letter, we explore conventional neck networks by analyzing information bottlenecks. We propose an enhanced network to address deficiency issue current networks. Our proposed network, which serves bridge between backbone head comprises global semantic (GSNet) feature fusion refinement module (FRM). The GSNet designed perceive contextual surroundings propagate discriminative knowledge through bidirectional pattern. FRM developed exploit different levels features capture comprehensive location information. validate efficacy efficiency our approach experiments conducted on two challenging datasets, DOTA HRSC2016. method outperforms existing approaches terms accuracy complexity, demonstrating superiority method.