Dense Condition-Driven Diffusion Network for Infrared Small Target Detection

Abstract

Infrared small target detection (IRSTD) is important in military and civilian applications. In recent years, numerous methods based on convolutional neural networks (CNNs) have already been explored in the field of IRSTD. However, due to the mismatch between the network’s receptive field and the size of the target, conventional CNN-based methods struggle to fully differentiate between the background and the small target and are prone to losing the small target in deeper layers. A dense condition-driven diffusion network (DCDNet) based on the conditional diffusion model is proposed to address the IRSTD task. The diffusion model can easily fit the distribution of infrared background images, thereby isolating the small targets from the distribution. Extracted features from original images are used as conditions to guide the diffusion model in gradually transforming Gaussian noise into the target image. A dense conditioning module is introduced to provide richer guidance to the diffusion model. This module incorporates multiscale information from the conditional image into the diffusion model. Multiple samplings can reduce the amplitude of background noise to enhance the target. Comprehensive experiments performed on two public datasets demonstrate the proposed method’s effectiveness and superiority over other comparative methods in terms of probability of detection ( $P_{d}$ ), intersection over union (IoU), and signal-to-clutter ratio gain (SCRG).