LRFNet: Toward Accurate Infrared Small Target Segmentation via Large Effective Receptive Field With Contextual Cues

Infrared small target segmentation is crucial for accurately identifying and delineating small, low-contrast targets in complex backgrounds. Due to the inherent limitations of infrared imaging sensors, targets often exhibit small volumes and blurred features, which pose difficulties for the accurate segmentation of infrared small targets. Significant advancements in the segmentation accuracy of infrared small targets have been achieved in recent years. However, existing methods mainly focus on exploiting the intrinsic characteristics of infrared small targets without utilizing the contextual information within the receptive field. In this article, we leverage a large kernel strategy and designed a novel network called large receptive field net, in which contextual cues are incorporated to assist to segment the infrared small targets. We propose a feature extraction block called context involved convolution layer (CICLayer) with more effectiveness compared with existing works. Integrated with a large convolutional kernel CICLayer is capable of excavating more semantic clues from larger effective receptive fields and discern finer details in infrared imagery. Furthermore, we design a dual-layer feature pyramid network to facilitate the superior feature fusion. Extensive experimental results on the IRSTD-1 k dataset and the NUDT-SIRST dataset demonstrate that the proposed method achieved 71.81% and 94.82% in <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">mIoU</i>, respectively. It outperforms the best method in existing works by approximately 6.29%, and 2.68%, respectively.