Renormalized Connection for Scale-Preferred Object Detection in Satellite Imagery

Abstract

Satellite imagery, due to its long-range imaging, brings with it a variety of\nscale-preferred tasks, such as the detection of tiny/small objects, making the\nprecise localization and detection of small objects of interest a challenging\ntask. In this article, we design a Knowledge Discovery Network (KDN) to\nimplement the renormalization group theory in terms of efficient feature\nextraction. Renormalized connection (RC) on the KDN enables ``synergistic\nfocusing'' of multi-scale features. Based on our observations of KDN, we\nabstract a class of RCs with different connection strengths, called n21C, and\ngeneralize it to FPN-based multi-branch detectors. In a series of FPN\nexperiments on the scale-preferred tasks, we found that the\n``divide-and-conquer'' idea of FPN severely hampers the detector's learning in\nthe right direction due to the large number of large-scale negative samples and\ninterference from background noise. Moreover, these negative samples cannot be\neliminated by the focal loss function. The RCs extends the multi-level\nfeature's ``divide-and-conquer'' mechanism of the FPN-based detectors to a wide\nrange of scale-preferred tasks, and enables synergistic effects of multi-level\nfeatures on the specific learning goal. In addition, interference activations\nin two aspects are greatly reduced and the detector learns in a more correct\ndirection. Extensive experiments of 17 well-designed detection architectures\nembedded with n21s on five different levels of scale-preferred tasks validate\nthe effectiveness and efficiency of the RCs. Especially the simplest linear\nform of RC, E421C performs well in all tasks and it satisfies the scaling\nproperty of RGT. We hope that our approach will transfer a large number of\nwell-designed detectors from the computer vision community to the remote\nsensing community.\n