A Rapid One-Stage End to End Hyperspectral Target Detection Model

Abstract

AbstractA hyperspectral target detection (HTD) task is essentially a binary classification task focusing on distinguishing specific targets from various backgrounds. However, most HTD methods consist of two stages, i.e. calculating a similarity score map and then producing the final detection map with a selected threshold, which cause cumulative errors and may influence the detection accuracy. In this paper, inspired by the risk estimation strategy and patch-free framework in hyperspectral classification tasks, a rapid one-stage end to end HTD model is proposed, which only makes use of target pixel samples to set up special binary classification tasks under HTD to avoid cumulative errors and constructs an end to end patch-free network to make use of the full image. Extensive experiments were made on three benchmark datasets and the experimental results indicate that our model can achieve superior performances in HTD.KeywordsTarget detectionOne-stage approachRapid inference speed