Classification of crispness of food materials by deep neural networks.

Abstract

Crispness is a textural characteristic that influences consumer choices, requiring a comprehensive understanding for product customization. Previous studies employing neural networks focused on acquiring audio through mechanical crushing of crispy samples. This research investigates the representation of crispy sound in time intervals and frequency domains, identifying key parameters to distinguish different foods. Two machine learning architectures, multi-layer perceptron (MLP) and residual neural network (ResNet), were used to analyze mel frequency cepstral coefficients (MFCC) and discrete Fourier transform (DFT) data, respectively. The models achieved over 95% accuracy "in-sample" successfully classifying fried chicken, potato chips, and toast using randomly extracted audio from ASMR videos. The MLP (MFCC) model demonstrated superior robustness compared to ResNet and predicted external inputs, such as freshly toasted bread acquired by a microphone or ASMR audio of toast in milk. In contrast, the ResNet model proved to be more responsive to variations in DFT spectrum and unable to predict the similarity of external audio sources, making it useful for classifying pretrained "in-samples". These findings are useful for classifying crispness among individual food sources. Additionally, the study explores the promising utilization of ASMR audio from Internet platforms to pretrain artificial neural network models, expanding the dataset for investigating the texture of crispy foods.