Augmenting Prototype Network with TransMix for Few-shot Hyperspectral Image Classification

• DOI: 10.48550/arxiv.2401.11724
• Publication Date: 2024-01-01
• Authors: Chun Liu; Longwei Yang; Dongmei Dong; Li Zheng; Wei Yang; Zhigang Han; Jiayao Wang
• Keywords: Robustness

Few-shot hyperspectral image classification aims to identify the classes of each pixel in images by only marking few these pixels. And order obtain spatial-spectral joint features pixel, fixed-size patches centering around are often used for classification. However, observing results existing methods, we found that boundary corresponding pixels which located at objects images, hard classify. These patchs mixed with multi-class spectral information. Inspired this, propose augment prototype network TransMix few-shot hyperspectrial classification(APNT). While taking as backbone, it adopts transformer feature extractor learn pixel-to-pixel relation and pay different attentions At same time, instead directly using cut from training, randomly mixs up two imitate uses synthetic train model, aim enlarge number training samples enhance their diversity. following data agumentation technique TransMix, attention returned is also mix labels generate better patches. Compared proposed method has demonstrated sate art performance robustness our experiments.