Standardization of Residuals for Detection of Unknown Target Classification of SAR images

The recent advancements in deep learning architectures have enabled the efficient and accurate classification of pre-trained targets in SAR images. However, encountering unknown targets during testing scenarios is unavoidable, causing misclassification and subsequently compromising classifier accuracy. Over the past few decades, a variety of approaches have been developed to address this challenge, yet the detection of unknown targets shows unstable fluctuations of accuracy dependent on the decision threshold value and the number of unknown targets. In this context, we propose a standardized method that demonstrates a relatively stable detection rate across multiple unknown target conditions concerning threshold metrics. As an example, we demonstrate feature extraction from the latent space of a deep learning network. Using these feature representations, residuals are computed from training data samples and then fitted with a Gaussian distribution to estimate statistical parameters. Anderson-Darling statistics are then applied to delineate the decision boundary for unknown targets. Furthermore, the comparison of histogram percentiles between known and unknown target classes is taken into account for establishing the fine-tuned statistical threshold for filtering unknown target samples. This systematic process consistently filters out unknown target classes across varying quantities of unknown classes at a fixed threshold value. Finally, we argue the practical significance of our approach within realistic scenarios.