Open-Set Classification of Maritime Radar Small Targets Using OpenMax Layer and Supervised Variational Autoencoder

Open-set classification of small targets in high-resolution maritime radars is a difficult problem. Due to various small targets, classifiers in service always encounter samples from unknown classes, which do not participate in training. Closed-set classifiers cannot distinguish between targets of known classes and unknown classes. In this paper, an open-set classification method using standardized time-frequency distribution images is proposed to classify small targets. It combines the existing closed-set ResNet classifier with a proposed supervised variational autoencoder for feature extraction and an improved OpenMax layer. To meet the open-set classification demands, a supervised variational autoencoder feature extractor is constructed by imposing the loss terms on the loss function to increase interclass distances, decrease intraclass distances of known classes, and constrain them to the surrounding region in the feature space, thereby mapping the standardized time-frequency distribution images into low-dimensional feature vectors. In the low-dimensional feature space, each known class is modelled by Weibull extreme distribution and Mahalanobis distance. The SoftMax score from the closed-set ResNet classifier and statistical models of the known classes in the feature space are combined into the OpenMax layer with rejection ability to samples from unknown classes. The proposed open set classification method is verified by maritime radar small target classification datasets from three open radar databases and island-based radar measured datasets. The experimental results show that the well-trained classifiers can effectively classify known classes of small targets and reject the samples from unknown classes.