Toward Detection of Small Objects Using Deep Learning Methods: A Review

Abstract

The field of computer vision, particularly object detection, has undergone significant changes. Most cutting-edge object detectors can accurately detect medium and large objects. Small object detection remains challenging for the majority of object detectors due to low resolution, lack of feature information, small objects appearing in unexpected areas or overlapping with other objects, and small object dataset limitations. Several solutions have been developed to address this issue. This paper provides a brief description and analysis of contemporary general object detectors, such as Faster R-CNN, SSD, and YOLO. In addition, we investigate several techniques to improve object detection performance, particularly for small object detection, from three perspectives: network improvement (multiscale feature, contextual information), input data optimization (super-resolution, image tiling), and dataset enhancement (data augmentation, creating own dataset). Implementing these techniques has been shown to improve the accuracy of contemporary object detectors, particularly for small objects.