A Multiscale Information Fusion Network Based on PixelShuffle Integrated With YOLO for Aerial Remote Sensing Object Detection

Deep learning-based object detection has made tremendous progress in the detection of aerial remote sensing targets. However, the issue of similar targets and multi-scale targets still becomes an obstacle in improving the detection accuracy. To address this issue, a multi-scale information fusion network based on PixelShuffle integrated with YOLO(MPS-YOLO) is proposed. First, to reduce the loss of deep semantic feature information of similar targets in the process of feature fusion, the feature pyramid network based on PixelShuffle(FPN-P) is introduced. Second, aiming at the phenomenon that gets stuck in identifying multi-scale targets, a multi-scale receptive field (MRF) module is designed to fuse the multi-scale information of the feature layer. Finally, to further enhance the detection result, an extra shallow feature map (ESF) is brought in to enrich the context information. Numerical results in public aerial remote sensing datasets show that, the proposed algorithm enhances the detection accuracy by 4.15% and has preeminent robustness to difficult-to-identify targets.