BiMA: Bidimensional multi-level attention embedded network for single-frame infrared small target detection

Abstract

As the depth of the detection network increases, features of small targets become less pronounced, and the model may inadvertently favor background clutter, thereby reducing the efficiency of target detection. Hence, a bidimensional multilevel attention-embedded network, named BiMA, is designed to simultaneously tackle original images rich in detail and filtered images that remove superfluous non-target elements, serving as its dual-input framework. Subsequently, a multi-scale fusion is raised to harness multi-scale properties of these images, offering a thorough grasp of both high-level and low-level features. In addition, a multi-level attention is integrated to further improve the detection capability. Extensive qualitative and quantitative experiments prove that BiMA not only surpasses contemporary approaches in robustly detecting small targets across a range of challenging scenarios but also has superior performance, e.g., higher probabilities of detection, reduced false alarm rates, and broader areas under receiver operating characteristic curves.