Feature Pyramid Full Granularity Attention Network for Object Detection in Remote Sensing Imagery

With the rapid advancement of deep learning, particularly the emergence of attention mechanisms applied to convolutional neural networks (CNNs), object detection in high-resolution remote sensing images has seen significant progress. However, due to the CNNs' inability to capture long-range dependencies and the high computational cost of the attention mechanism, object detection in remote sensing images remains a challenging task. To address these issues, this paper introduces a novel feature pyramid full granularity attention module (FPFGAM) designed to learn long-range dependencies, dynamically attend to strongly correlated features, and reduce GPU memory overhead. Initially, we perform adaptive filtering of feature regions at the coarse-grained level. This process reduces the computational burden caused by weakly correlated features. Subsequently, we perform fine-grained pixel-level queries on several strongly correlated regions to enhance long-range dependent feature learning. We propose a feature pyramid full granularity attention network (FPFGANet) by embedding the feature pyramid full granularity attention module into the backbone network ResNet50 and the feature pyramid network (FPN). FPFGAM can be easily inserted into different layers to improve object detection accuracy in remote sensing images. Finally, we evaluate our method on three commonly used public remote sensing object detection datasets: NWPU VHR-10 and DIOR. The empirical results confirm the effectiveness of our approach.