DST-HRNet: Infrared dim and small target extraction algorithm based on improved HRNet

Abstract

Single frame infrared dim and small target detection is always a difficult subject due to the lack of obvious target features and the difficulty of target extraction.In this paper, with the single frame infrared image dataset as the data source, and dim and small targets as the research object, infrared dim and small targets are extracted by separating the targets from the background. According to the task requirements, this paper proposes an infrared dim and small target extraction algorithm based on the improved HRNet. Based on HRNet, a semantic segmentation network, the algorithm optimizes the processing flow by introducing the attention mechanism module, so as to effectively extract the image surface features and improve the detection precision. In this paper, ablation experiments are conducted in detail in the single frame infrared small target (SIRST) dataset. A comparison is made of the effect of each attention mechanism module added to different positions of the network in HRNet Among them, when SE module (an attention mechanism module) is added to the first two steps of down-sampling in HRNet, the enhanced effect is most obvious, with the value of IoU reaching 76.9%. The experimental results show that the algorithm can be effective in detecting single frame infrared dim and small targets.