Overhead Object Detection with Channel Attention for High-Resolution Multi-Spectral Satellite Imagery and DMP-extracted Shape Features

Deep convolutional neural networks (DCNNs) have shown excellent performance in object detection tasks over the last five years. The widespread adoption of DCNNs in computer vision (CV) applications is seen in many domains, including overhead remote sensing (RS), where the adoption of DCNNs has enabled robust CV applications ranging from broad area search to scene classification. While traditional CV algorithms and applications are developed utilizing three-channel, RGB tensors as input, many overhead remote sensing applications have additional channels available in the form of multi-spectral imagery, as well as from techniques that generate additional input channels as a preprocessing step to feature extraction. However, the additional channels available to DCNNs have different levels of utility for object detection tasks, even at the dataset level. To that end, here we propose utilizing channel attention blocks to enable DCNNs to learn the relative importance of each input channel to the network. We then apply our novel architecture to the xView challenge dataset and show increased performance in both AP50 and AR50.