An Enhanced Algorithm for Small Object Detection based on Thermal Imaging Using YOLOv8-EPB

Abstract

Object detection is one of the most important and challenging problems in the computer vision domain. Using the power of deep models, researchers have carefully explored and made significant contributions to increasing the effectiveness of object detection and related tasks, such as object identification, localization, and segmentation. This progress is due to the rapid progress of deep learning in the past decade. However, object detection in thermal imaging has certain challenges and has potential uses in areas like autonomous driving, security, and surveillance. When applying several popular object detection algorithms to ground-based thermal imaging, the main obstacles include the small size of the targeted object, low-quality images, obstruction, and varying illuminating conditions. In this study, to address this problem enhanced version of YOLOv8 termed asYOLOv8-EPB algorithm has been proposed to target small-size objects in ground-based thermal images. Initially replacing the CSPDarknet53 backbone with EfficientNet-B4 reduces model parameter's computational complexity and increases inference speed. In addition, a new compact target-detecting layer and head have been created to reduce noise in thermal imaging. Lastly, adding a Bidirectional Feature Pyramid Network (BiFPN) to the neck section improves model generalization by lowering detection errors caused by scale deviations and complex situations. The study evaluates a proposed algorithm through ablation experiments and comparisons with other algorithms, focusing on detection performance. The algorithm obtained a mean Average Precision of 92.3% in a self-made thermal imaging dataset, with an accuracy increase of 4.7% compared to regular YOLOv8 models and outperforming other leading-edge detection algorithms.