An enhanced convolutional neural network architecture for nondestructive detection of microbial contamination on eggshells through hyperspectral imaging.

Eggs, a key dietary staple globally, are valued for their high nutritional content. However, the safety of eggshells is often compromised by microbial contamination. In this study, we developed a reliable, noninvasive method for detecting the aerobic plate count on eggshells through hyperspectral imaging (HSI) combined with a deep learning approach. Visible-near infrared HSI data (450-1100 nm) were collected for 108 egg samples. To address the high dimensionality of HSI data, three wavelength selection methods-competitive adaptive reweighted sampling, selectivity ratio, and variable importance in projection (VIP)-were used to extract the most informative spectral features. A modified convolutional neural network (CNN) enhanced with channel attention (CA) and depthwise separable convolution (DSC; termed CA-DSC-CNN) was developed to efficiently model spectral-spatial features while reducing computational complexity. CA-DSC-CNN with wavelengths selected using VIP outperformed its state-of-the-art counterparts, yielding a correlation coefficient of 0.8959 for the prediction set and a root mean square error of 0.2396. These values surpassed those of conventional chemometric models, including partial least squares regression, multiple linear regression, support vector regression and standard CNNs. This study demonstrated that HSI integrated with advanced deep learning techniques enabled the rapid, nondestructive detection of microbial contamination on eggshells, contributing to improved safety in egg processing and storage.