Deep Spectral-spatial Features of Snapshot Hyperspectral Images for Red-meat Classification

摘要

We investigate the potential and accuracy of snapshot hyperspectral imaging for authentication and classification of red-meat species. Snapshot hyperspectral images are acquired of lamb, beef, and pork samples. We consider 13 muscles types of standard loin and leg chops. We propose a deep 3D convolution neural network (CNN) architecture for extracting and classifying spectral-spatial learned features of red-meat. We present a comparison with state-of-the-art models including partial least-square discriminant analysis and support vector machines. Our results show that the proposed 3D-CNN model outperforms the state-of-the-art models with 95.81% and 96.01% for overall accuracy and average F₁ score, respectively. Visualization results show that the proposed 3D-CNN model is able to convert snapshot hyperspectral image data into an intelligent representation with accurate separation between red-meat types. This study opens the door for more research towards real-time and completely portable hyperspectral imaging systems due to the ability of snapshot hyperspectral cameras to work at video rate.