Confidence Fusion With Representation Distribution and Mixture of Experts for Multimodal Radar Target Recognition

Radar target recognition technology based on multimodal fusion provides a comprehensive target description that encompasses not only the target's echo structure but also its physical motion characteristics or other attributes, thereby enhancing recognition performance. However, in practical applications, it faces the challenge of uncertainty fusion due to variations in modal data quality. In this paper, we delve into the issues related to uncertainty in radar target recognition and propose a confidence fusion method with representation distribution and mixture of experts (CFRDMoE). This method comprises three primary modules: representation distribution modeling, singlemodal mixture of experts learner, and confidence fusion. Firstly, the representation distribution module modeling extends traditional feature point embedding to distributional embedding and obtains modality-specific tokens. It utilizes the mean to represent the central location of features in the latent space and the variance to describe feature disturbances. Secondly, the single-modal mixture of experts learner module adaptively selects the appropriate feedforward neural network (FFN) through a router from a pool of experts composed of multiple FFNs for feature learning. Lastly, the confidence fusion module calculates modal reliability based on the single-modal evidence predicted by the network. We perform the weighted fusion of single-modal results based on relative confidence values to improve the quality of fusion. Experiments are conducted with various configurations on a measured dataset, and the results demonstrate superior recognition performance under both normal and modal disturbance conditions.