📝 Research Experience

Open set recognition based on a hyperspherical embedding network

• Model extracted features onto a unit hypersphere with von Mises-Fisher (vMF) distribution to improve the inter-class separability of feature representations.

• Measure the similarity of samples to each class center and design a discrimination criterion to detect unknown targets based on this metric.

• Calibrate the predicted probability based on uncertainty estimation.

• Implement the algorithm and evaluate the performance on two public satellite imagery datasets (MSTAR).

A threshold-free open-set recognition network for satellite imagery

• Propose a threshold-free open-set learning network for satellite target recognition.

• Improve the network structure of GAN to formulate the open-set recognition problem as a K+1 classification task.

• Implement the algorithm, evaluate the performance, and compare it with four SOTA models on two public satellite datasets (MSTAR and SAMPLE).