Bounding Box Regression Network for Infrared Small Target Detection With Adaptive Receptive Field and Cross-Scale Fusion

Single-frame infrared small target (SIRST) detection is crucial for both military and civilian applications, but remains challenging due to low resolution and small target sizes. Most existing methods model the detection task as a semantic segmentation task, which requires high-resolution feature maps and incurs significant computational costs. Moreover, manual annotations often struggle to achieve pixel-level precision, and the inherent ambiguity in the annotations can affect the training outcomes. This article treats the SIRST detection task as a bounding box regression problem and proposes a novel target detection network architecture, named adaptive fusion bounding box regression network (ABRNet). Specifically, to address the challenges posed by complex and changeable backgrounds, we design an adaptive receptive field (ARF) module. This module utilizes spatial selection masks to choose feature maps with varying receptive field (RF) sizes, thereby leveraging the unique prior knowledge inherent in different scenarios. In addition, we introduce a cross-scale feature encoding fusion (CEF) structure to alleviate the network's low tolerance to bounding box perturbations. The module fuses multiscale local and global features to capture the fine details of small targets. By combining high-dimensional features with detailed features, it facilitates accurate bounding box regression, thereby improving detection performance. Additionally, we employ linear interval mapping to achieve dynamic balancing of hard samples. Experimental results on public datasets demonstrate ABRNet's superiority over state-of-the-art (SOTA) methods.