Multi-Scale Object Detection in Aerial Images Based on Cascade R-CNN

Abstract

Computer vision is widely used in urban management, environmental protection, and intelligent monitoring. One of the issues is that in dense crowds, it is difficult to accurately identify objects of different sizes and angles. To address the aforementioned shortcomings, this paper proposes a detection framework based on an improved Cascade R-CNN. This will allow for the identification of multi-scale targets in aerial images. The framework of the proposed method consists of high-level multi-scale feature aggregation, scale-aware loss functions, and multi-branch context refinement modules. Conduct numerous practical experiments using typical data from various real-life scenarios to determine whether the aforementioned framework is effective in all cases. The results show that this new method outperforms traditional methods in terms of average precision for small and overlapping targets, and remains effective in many cases. According to the comparison and ablation experiments, all modules need to improve detection accuracy and reduce false positives. Due to its unique structure, this system is highly adaptable to high-density and highly variable remote sensing environments. This paper provides practical support for the automated analysis of large-scale and reliable remote sensing data, offering high-performance aerial image target detection as a foundation for efficient and scalable technology.