IIANet: Information Interactivity Attention Network with adversarial learning for infrared small object detection

Infrared small object detection plays an important role in maritime surveillance, early warning systems, and precise positioning. However, existing deep learning methods lack attention to local information and overlook the interactivity between local and global information. To solve this problem and explore a more effective feature extraction network for infrared small object detection, we propose an information interactivity attention network (IIANet), which decouples the optimization of miss detection and false alarm by adversarial learning. Specifically, the information interactivity attention is achieved by the combination of the global spatial attention module and pixel-wise local attention module where they can complement each other to generate finer attention. Correspondingly, to make full use of the extracted interactive features, a more sophisticated information aggregation discriminator is designed by using the intensive skip-addition module which further realizes information communication among different semantic information. Finally, we apply the pixel equalization strategy to further promote the learning ability of the data-oriented model by enhancing the visual contrast between the objects and the background. Extensive experiments and visualization results demonstrate that our approach can achieve state-of-the-art performance with a higher detection rate and lower false alarm rate. Besides, ablation experiments verify the effectiveness of each module.