MI2T-UNet: Integrating Multi-Scale Input Image and Top-hat Context for Infrared Small Object Segmentation

Abstract

The segmentation of small objects in the infrared spectrum, commonly referred to as Infrared Small Object Segmentation (ISOS), holds significant importance in many applications, including but not limited to target detection, surveillance, and autonomous navigation. The task of achieving precise image segmentation of objects of interest is formidable, owing to the diminutive size of the objects in question, the intricate nature of the background, and the limited contrast differential. This research introduces a novel mobile-UNet architecture with attention gates, namely MI2T-UNet, which leverages multi-scale input images and top-hat operation. The proposed methodology involves the implementation of architectural modifications to improve the scale adaptation of the model. Specifically, these modifications include concatenating down-scaled input images with corresponding encoder layers, utilizing multi-kernel top-hats in the encoders and attention gates, and integrating encoder outputs with subsequent encoder layers to facilitate progressive information flow. This article also compares the proposed architecture against the top-performing models on the SIRST and IRSTD datasets, utilizing metrics for pixel level such as IoU/NIoU, object level including Pd/Fa, and model level comprising F-Score/AUC. The experimental results demonstrate that MI2T-UNet outperforms state-of-the-art models in terms of segmentation, with an average margin of over 1±0.25% for evaluation metrics. This research contributes to the ISOS field by proposing a novel approach to enhance the discriminating ability of the network. Specifically, a modified mobile-UNet has been developed, tailored to the ISOS requirements. The proposed approach has the potential to make significant strides in infrared image analysis and streamline various applications that hinge on precise ISOS.