[Paper] Evaluation of Deep Feature Reconstruction for Pixel-Wise Hyperspectral Image Classification

抄録

Most state-of-the-art methods for pixel-wise hyperspectral image (HSI) classification are based on the Convolutional Neural Network (CNN). In this paper, we introduce a feature reconstruction module (FRM) into the CNN-based network of pixel-wise HSI classification to improve classification accuracy. FRM can extract essential characteristics in the original matrix of CNN features by low-rank approximation using matrix factorization. We compare the classification accuracy before and after the introduction of FRM into the CNN-based network of pixel-wise HSI classification to validate its effectiveness. Experimental results demonstrate this method improved classification accuracy. We also visualized and compared the original CNN features and the reconstructed CNN features to evaluate which features contributed to the improvement in classification accuracy.