A Review of Advances in Deep Learning-Based Small Object Detection

Small object detection (SOD), as a crucial component of object detection, frequently appears in aerial remote sensing detection and autopilot. However, SOD has proven to be a challenging task due to the low resolution. This review briefly introduces the definition of small objects and the related datasets for specific tasks and then offers a thorough analysis of the research progress of deep learning-based SOD algorithms in terms of techniques to enhance detection performance, including data augmentation, super-resolution, context-based information, multi-scale representation, anchor mechanism, and sample assignment strategies. Towards the end, several research directions that may improve the detection performance are proposed by summarizing and analyzing the current shortcomings and challenges of SOD.