Feature Enhancement Network for Object Detection in Optical Remote Sensing Images

Automatic and robust object detection in remote sensing images is of vital significance real-world applications such as land resource management disaster rescue. However, poor performance arises when the state-of-the-art natural image algorithms are directly applied to images, which largely results from variations scale, aspect ratio, indistinguishable appearances, complex background scenario. In this paper, we propose a novel Feature Enhancement Network (FENet) for optical consists Dual Attention (DAFE) module Context (CFE) module. Specifically, DAFE introduced highlight network focus on distinctive features objects interest suppress useless ones by jointly recalibrating spatial channel feature responses. The CFE designed capture global context cues selectively strengthen class-aware leveraging image-level contextual information that indicates presence or absence classes. To end, employ encoding loss regularize model training promotes detector understand scene better narrows probable categories prediction. We achieve our proposed FENet unifying into framework Faster R-CNN. experiments, evaluate method two large-scale datasets including DIOR DOTA demonstrate its effectiveness compared with baseline methods.