Evaluation of sensory crispness of dry crispy foods by convolutional neural networks

• Ano de defesa: 2023
• Autor(a) principal: Lopes, Rafael Zinni
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: eng
• Instituição de defesa: Biblioteca Digitais de Teses e Dissertações da USP
• Palavras-chave em Português: Convolutional neural network; Crispness; Crocância; Food materials; Librosa; Materiais alimentícios; Rede neural convolucional; Toast; Torrada
• Link de acesso: https://www.teses.usp.br/teses/disponiveis/74/74133/tde-09022024-105654/

Convective drying is traditionally used to dehydrate food, reducing volume and water activity for easy transportation and storage. During drying, foods undergo volume reduction due to moisture loss, resulting in changes in the solid matrix and the formation of a crispy structure when crushed or fractured. This study focused on developing methods for quantifying and classifying crispy dried foods, such as potato chips, toasts, and fried foods like french fries and fried chicken, which were investigated. Compression profiles and sound noise were determined using a lever device covered by a noise suppression box. The captured sound was transformed into different parameters using Python and Mathematica Wolfram libraries. The power spectrum of the sound signal was obtained using the discrete Fourier transform method in Wolfram, while Onset Strength and Mel Frequency Cepstral Coefficients (MFCC) were obtained using the Librosa library. The sound spectra, Onset Strength, and MFCC were processed using neural networks to classify the crispness of fried chicken, potato chips, and toasts. The classification models using DFT and MFCC signals achieved an accuracy of over 95%. This study allowed the description of crispy sounds based on the intensity and duration of the signal. A second study utilized Python code and the Librosa library in an attempt to generate a dimensionless number, called the Zeta value, for classifying crispness intensity. The Zeta value was calculated based on Root Mean Squared Energy values multiplied by peak intensities within 1-second intervals. Experimental validation of the Zeta value was performed by acquiring crispness noises for toasts and French fries while monitoring moisture and storage time. Zeta behavior aligned with the crispness behavior in the tests of increasing and decreasing crispness over time.