APTNet: Adaptive Partial Transformer Network for Infrared Small Target Detection

Infrared small target detection (IRSTD) plays a vital role in various applications such as surveillance and rescue missions. However, existing methods struggle to effectively integrate contextual information for small targets, with background noise and limited long-distance dependency learning further constraining detection performance. To address these issues, this paper proposes a novel adaptive partial transformer network based on asymmetric U-Net architecture, namely APTNet, which enhances contextual information integration and accurately detects small objects in complex scenes, improving detection performance. Specifically, a Double Residual Attention Block (DRAB) is designed to enhance the contrast of dim targets within complex backgrounds through channel and spatial attention mechanisms, thereby improving discrimination capability for small targets. Furthermore, leveraging the context pooling connections and DRAB output features, an Adaptive Partial Transformer Block is proposed. This block first proposes an adaptive channel splitting factor, achieving contextual information integration and long-distance dependency learning with low computational overhead. Extensive experiments on two public datasets demonstrate that the proposed method significantly outperforms the existing state-of-the-art IRSTD methods in terms of intersection over union (IoU), probability of detection (Pd), and false alarm rate (Fa). The implementation of our code can be accessed at https://github.com/Wangtao-Bao/APTNet.