Enhanced Small Object Detection in Remotely Sensed Images using Hybrid ResNet50 and Feature Pyramid Network

Developing automated systems for detecting small objects in aerial images is a challenging task due to the complexity of the scenes and the minute size of the objects. The problem is critical for applications in surveillance, tracking movements on sea and land borders, wildlife monitoring, disaster management and urban planning like infrastructure analysis, traffic management, green space management, etc. However this automated detection is hindered due to the tiny-sized objects, occlusions, camouflages, noise clutters, minimal pixel information, etc. To address these challenges, the work proposes a deep learning approach utilizing ResNet50 integrated with a Feature Pyramid Network (FPN). The main goal of this approach is to improve detection accuracy and robustness for small objects amidst complex aerial scenes. The work presents a comprehensive analysis covering data preprocessing, model architecture, training, and performance evaluation. Notably, the ResNet50-FPN model exhibits superior performance with an accuracy, precision, recall, and F1-score of 84.12%, 84.52%, 85.69%, and 85.10% respectively. Using the SODA-A dataset tailored for small object detection, we provide detailed insights into the implementation and comparative analysis of this model, highlighting its effectiveness in small object detection tasks and illustrating its practical utility in diverse real-world scenarios.