Dictionary cache transformer for hyperspectral image classification

doi:10.1007/s10489-023-04934-5

Dictionary cache transformer for hyperspectral image classification

2023-08-29

3

PRE

AI

DOI：10.1007/s10489-023-04934-5

Original

Original request for help

Save

Abstract

En 中文

The spectral anomalies, limited training samples, and noisy training labels pose significant challenges to accurately classifying hyperspectral images (HSIs). To address these issues, we propose a novel dictionary cache transformer (DiCT) for HSIs classification, leveraging a combination of a group self-attention mechanism and a dictionary cache module. Specifically, the group self-attention mechanism binds the features local to the pixel into a group, thus alleviating the interference caused by pixel spectral anomalies. Furthermore, we capture representative structural information from different samples using their discriminative features to construct the dictionary cache module. The dictionary cache module enhances features by fusing sample features and the most similar element in the dictionary cache, thus improving the model's resilience to noisy training labels. Experiments on five HSIs datasets demonstrate the proposed DiCT's superiority in classification performance and robustness to noisy training labels.

Keywords:

Vision transformer

Dictionary cache

Hyperspectral image classification

Noisy training labels

Feature enhancement

Dictionary cache transformer for hyperspectral image classification

Abstract

Journal

Organization