Tiny Object Detection Enhancement for Large-Scale Remote Sensing Imagery

The effective detection of targets of interest in large-scale remote sensing imagery has always been a crucial application task. Detecting tiny objects in these scenes poses particular challenges, as it can lead to the loss of robust features and difficulties in generating positive sample matches. To address this application scenario, this paper proposes a method that embeds the use of vague search (VGS) to match high-potential candidate regions within a coarse-to-fine detection framework. Additionally, a new centroid distance measurement (CDM) is proposed to distinguish between positive and negative proposals. The combination of VGS and CDM can effectively increases the number of positive samples that can be refined in the high-dimensional detection layer, thereby improving the precision of detecting tiny objects in large-scale scenes. We tested and validated our model on a high-resolution optical remote sensing scene dataset, where the overall mean Average Precision (mAP) for the target of interest, bridges, improved from 56.2% to 71.3%, achieving impressive results.