Mondrian conformal prediction enhanced LSTM for birds and drones recognition
Nannan Zhu, Zepu Xi, Hongbo Chen, Shiyou Xu, Yifan Wang, Fuli Zhong · IET conference proceedings. · 2024
Deep learning has achieved remarkable success in target recognition. However, its efficacy has raised distrust within radar research due to challenges in interpretability and generalization. This study introduces the Mondrian conformal prediction enhanced long short-term memory (MCPLSTM) framework, a pioneering fusion of conformal prediction and deep learning. Applied to the recognition of birds and drones in radar sequences, the MCPLSTM framework offers not only classification results but also the conferral of confidence and reliability for each individual classification. The integration of confidenc e levels into the recognition process empowers precision control, enabling the interpretation of deep learning outcomes through a statistical prism. Empirical validation is realized through extensive analysis of real-world data from an airport bird detection radar system, confirming the effectiveness of the proposed algorithm.