Detection of Early Browning in Pears Based on Time–Frequency Images of Vibro-Acoustic Signals and Improved MobileNetV3

Abstract

A method utilizing vibroacoustic nondestructive testing technology was proposed to address the difficulty of detecting early internal browning in pears. It combined the S-transform signal analysis method and MobileNetV3 neural network to distinguish pears with different degrees of browning. After converting the one-dimensional response signal into a two-dimensional time–frequency image through the S transform, these time–frequency images were then categorized into healthy, slight browning, and severe browning pears based on the actual degree of browning observed in the pears. These three classes of datasets were inputted into an improved MobileNetV3 model, which incorporates a random cropping module (RC) and attention mechanism (CA). The constructed MobileNetV3-RA model achieved an overall classification accuracy of 96.79%. Specifically, the accuracy values were as follows: 97.50% for healthy pears, 93.75% for slightly browning pears, and 95.83% for severely browning pears. Notably, on the imbalanced test set, the F1, Ka, MCC, and IAM values for the MobileNetV3-RA model were 94.05%, 91.72%, 91.57%, and 83.97%, respectively. Additionally, MobileNetV3-RA has faster detection speed and fewer parameters, with respective values of 0.27s and 4.45M, compared to other common classification models. Therefore, this study provides an effective nondestructive testing method for early internal browning of pears and provides some research ideas and methods for internal quality detection of other fruits and vegetables.

Graphical Abstract