Moldy peanuts identification based on hyperspectral images and Point-centered convolutional neural network combined with embedded feature selection

Abstract

Peanuts are prone to mold when stored improperly, resulting in the production of aflatoxin, which is a potential carcinogen. Near-infrared hyperspectral imaging is an emerging technique for the nondestructive detection of food contamination. In this study, hyperspectral images (HSIs) of healthy and moldy peanuts were acquired for classification. A point-centered convolutional neural network combined with embedded feature selection (PCNN-FS) was proposed, which integrates feature selection, feature extraction, and classification into an end-to-end trainable network. Conventional feature selection and classification algorithms were compared with the proposed method. The lightweight point-centered convolutional neural network (PCNN) based on five key bands (988, 1321, 1933, 2234, and 2563 nm) selected by the PCNN-FS achieved an accuracy of 97.98%, demonstrating a competitive performance against that based on full bands (98.51%). Compared with conventional methods, the lightweight PCNN performed better with fewer bands. The results indicate that the combination of HSIs and PCNN-FS has great potential for nondestructive identification of moldy peanuts.