Wei Hua, Qili Chen
A comprehensive survey paper that systematically organizes the main challenges and deep learning optimization strategies for small object detection in aerial images.
Small objects (e.g., small vehicles, buildings) in aerial remote sensing images have few pixels, complex backgrounds, and appear at various angles and scales, making detection very difficult. Directly applying deep learning detection techniques successful in natural images leads to significantly degraded performance.
The paper first summarizes the main challenges of small object detection (small size, viewpoint specificity, background complexity, scale and orientation diversity). Then, it categorizes and analyzes deep learning optimization strategies into data augmentation, network architecture improvement, loss function design, multi-scale processing, etc. It also introduces commonly used datasets and evaluation metrics, and compares experimental results of state-of-the-art detection algorithms.
It systematically compares the advantages and disadvantages of various optimization strategies and suggests future research directions (e.g., lightweight models, self-supervised learning, domain adaptation). This survey provides a useful reference for researchers in aerial image small object detection.