关键词: 超声成像技术；高光谱成像技术；调理牛排；数据融合

Abstract: A novel method to discriminate between raw and restructured beef steal was developed by using hyperspectral and ultrasonic imaging. Hyperspectral and ultrasonic images of samples were collected from which texture feature values were extracted by gray level co-occurrence matrix (GLCM). Four discriminant models were established separately using linear discriminant analysis, K-nearest neighbor (KNN), back propagation artificial neural network, and extreme learning machine (ELM) based on the hyperspectral image data, the ultrasonic image data or their fusion. We comparatively evaluated three variable selection techniques: successive projections algorithm (SPA), competitive adaptive reweighted sampling (CARS), and variables combination population analysis (VCPA). The results suggested that the texture of restructured beef steak was uniform, and the ultrasonic image signal was weak and uniform, which was distinct from that of raw beef steak. The best models for hyperspectral and ultrasound imaging were KNN and ELM, with prediction set identification rates of 95.00% and 90.00%, respectively. After data fusion, the prediction set identification rate of the best model ELM was 97.50%. The prediction set identification rates of the models established based on the texture variables selected by CARS and VCPA were 100.00%. The results showed that data fusion between hyperspectral and ultrasonic imaging combined with a suitable variable selection method can discriminate between raw and restructured beef steak quickly and accurately.

Key words: ultrasonic imaging technology; hyperspectral imaging technology; restructured steak; data fusion