Small Object Detection in Aerial Thermal Images Using the Hybrid ViT-CapNet Algorithm

Abstract

The automatic recognition of objects in aerial thermal images plays a very crucial role in many real-time applications. The advancement of Unmanned Aerial Vehicles (UAVs) has enabled real-time surveillance and monitoring in sectors such as defense, agriculture, and disaster management. Small target size, limited spatial resolution, low visibility under night-time conditions, background clutter, pose, and scale variations are significant challenges to the object detection techniques. This paper presents a unified architecture for the design and development of a Small Object Detection (SOD) system specifically optimized for mini UAV platforms. The proposed approach integrates a Grey Wolf Optimization (GWO)-based image enhancement subsystem with a hybrid Vision Transformer-Capsule Network (ViT-CapNet) deep learning model to significantly improve object detection accuracy in complex thermal environments. To evaluate the robustness of the proposed framework, it was tested through a comprehensive analysis involving leading deep learning models including CNN, FRCNN, COMNET, RESNET, SMPNET, and Vision Transformer (ViT) with Precision, Recall, F1-score, Map 0.5 (IoU), and Speed in FPS being considered as the evaluation metrics. The performance comparison of these algorithms clearly shown the ViT architecture offers superior performance for small object detection in UAV-based thermal imagery with the highest precision and F1-score.