A hyperspectral image classification framework and its application

• A new descriptor called LRP is proposed to describe subtle texture for 2D image.
• Texture enhancement is built to extend LRP to represent HSI.
• A supervised manifold learning method called DLPP is proposed to reduce dimension.
• The proposed framework outperforms other methods with experiments.

Classification is a critical step to make full use of the hyperspectral data. The most current approaches perform well for analyzing the macro texture, but they often fail to deal with the micro texture. Thus, this study proposes a general framework for the material with micro texture based on Hyperspectral Image (HSI) technique. In this framework, Local Response Pattern (LRP) is firstly proposed to describe 2D image texture to preserve more structural information and keep less sensitive to image conditions. Then, LRP is extended to represent HSI with Texture Enhancement (TE) by considering opponent relationships between pairs of bands. After that, Discriminated Locality Preserving Projection (DLPP) is proposed to reduce data dimension in a linearizing nonlinear manifold way. Finally, experiments on the hyperspectral images of fresh and frozen-thawed fish fillets are conducted. The results demonstrate that the proposed framework is efficient in terms of both recognition rates and robustness.