Semantic segmentation of infrared ships based on scene-aware priors

In the infrared ship target detection task, natural interference such as fish scale light, bright sea surface, clouds, and rainfall are prone to occur in the scene. Common detection algorithms focus on the target area and ignore the effective information in other areas of the image, so it is difficult to effectively deal with the interference of complex infrared scenes on the sea surface. In this paper, "semantic segmentation" is introduced into the field of infrared ship target detection, and the characteristics of global image processing of semantic segmentation are used to obtain deep semantic information of images. Combined with prior theory, Scene-Aware Prior semantic segmentation Network (SAPN) is proposed. In addition, a dataset of infrared ship target semantic segmentation is constructed and finely annotated for training and testing of the infrared ship semantic segmentation network. Through design experiments, SAPN can effectively deal with complex scenes. On the test set of infrared ship target semantic segmentation dataset, the mIoU of the model in this paper is improved to 70.26%, and it is better than other state-of-the-art networks. SAPN aims to effectively utilize the whole image information, not limited to a single target area. It can use all the information of the image as an auxiliary criterion for ship target recognition, and can provide an information basis for future intelligent decision-making and mission planning.