MOE-IR: Infrared Dim Small Target Detection Method With Mixture of Experts Feature Extraction

Infrared small target detection holds significant importance across various domains, including military and security applications. Nevertheless, detecting small targets is highly challenging due to their minimal pixel presence in images, indistinct features, and complex backgrounds. Existing detection methods are often interfered by complex backgrounds, resulting in unsatisfactory detection results. To overcome this challenge, this study introduces the mixture of experts infrared small target detection (MOE-IR) method. The core idea of this method is to construct a mixture of experts feature extraction network to perform rich feature extraction and complex background suppression on small targets, respectively, so as to achieve robust infrared small target detection in complex backgrounds. Specifically, the MOE-IR comprises a target feature extraction expert and a background suppression expert. The target feature extraction expert focuses on enhancing the features of infrared small targets, while the background suppression expert aims to mitigate background clutter. Additionally, an adaptive gate controlled network is incorporated to dynamically assign weights to the two experts based on the input infrared image, ensuring effective detection of infrared small targets across diverse and complex scenarios. Extensive experiments demonstrate that the proposed algorithm surpasses existing infrared small target detection methods in terms of detection accuracy and false alarm rate. It can reliably and stably identify small targets within infrared images, thus offering an effective solution for practical infrared small target detection applications.