SMALL OBJECT SEGMENTATION IN INFRARED IMAGES

Abstract

Infrared imaging technology offers unique capabilities for detecting objects in low-light or obscured conditions, making it valuable for applications like security, industrial inspection, and search and rescue operations. However, detecting and segmenting small objects in infrared images remains a challenging task due to low contrast, noise, and background clutter. This thesis proposes a novel deep learning methodology that addresses these challenges by incorporating three innovative components within a U-shaped neural network architecture: the Focus Block, the Fast Fourier Convolution (FFC) Block, and the Reasoner Block. The Focus Block mimics human visual perception by enhancing the contrast between small objects and their surroundings, effectively separating targets from complex backgrounds. The FFC Block leverages frequency domain analysis to distinguish between spatial information of small objects and noise, enabling effective noise suppression. The Reasoner Block, inspired by transformer architectures, models semantic and spatial relationships between image regions, improving the understanding of object context. By integrating these components, the proposed methodology aims to achieve robust and accurate small object segmentation in infrared images. Extensive experiments were conducted on two benchmark datasets: SIRST and IRSTD. The results demonstrate that the proposed approach outperforms existing state-of-the-art techniques, achieving higher intersection over union (IoU) scores and improved detection rates while reducing false alarms. This thesis presents a significant advancement in the field of infrared image analysis by addressing the critical challenge of small object segmentation, enabling more effective utilization of infrared imaging technology across various domains.