SMPISD-MTPNet: Scene Semantic Prior-Assisted Infrared Ship Detection Using Multitask Perception Networks

Chen Hu, Xiaogang Dong, Yian Huang, Lele Wang, Liang Xu, Tian Pu, Zhenming Peng

Abstract: Infrared ship detection (IRSD) is crucial for numerous applications but faces challenges, such as small targets and complex backgrounds, resulting in misdetections and false alarms. In order to address these challenges, we propose the scene semantic prior-assisted infrared ship detection using multitask perception network (SMPISD-MTPNet). This network employs multitask perception: one task is to predict targets, and the other focuses on scene perception to suppress false alarms caused by background interference. To highlight dim and small targets, we use the scene semantic extractor (SSE) to guide the network using features extracted based on expert knowledge and the gradient-based module to enhance the edge and point features. We apply data augmentation to the networks and employ a training trick called soft fine-tuning to improve the network’s generalization and suppress the distortion caused by the augmentation process. Due to the unavailability of datasets with appropriate scene labels for scene perception, we have developed a new dataset called the infrared ship dataset with scene segmentation (IRSDSS). In addition, we have enhanced an existing dataset by adding scene masks and created the enhanced infrared ship detection dataset (EISDD). Our evaluations using both IRSDSS and EISDD demonstrate that SMPISD-MTPNet exceeds contemporary state-of-the-art (SOTA) methods in accuracy. The source code and dataset for this research can be available at: https://github.com/greekinRoma/SMPISD-MTPNet.