In this study, hyperspectral imaging was used to visualize the water content and springiness of preserved egg gels and to predict different quality grades. First, the hyperspectral information of qualified and high quality preserved eggs was collected, and their water content and springiness were measured. Then, the original spectral data were transformed by Savitzky-Golay (S-G), first derivative (FD) or Savitzky-Golay and first derivative (S-G-FD) to analyze their correlation with water content and springiness values. We identified and excluded outliers by Monte Carlo-partial least squares (MCPLS), and partitioned the sample sets by sample set partitioning based on joint X-Y distance (SPXY). The characteristic wavelengths were selected using successive projection algorithm (SPA) and the uninformative variable elimination (UVE) method, and a multiple stepwise regression model (MSR) was established to predict the water content and springiness of preserved eggs. The results showed that UVE-MSR was the optimal model for predicting water content. Its determination coefficient and root-mean-square error (RMSE) were 0.882 and 0.583, respectively, and its relative percent deviation (RPD) was 2.1. The optimal model for predicting springiness was SPA-MSR, whose determination coefficient and RMSE were 0.903 and 0.348, respectively, and whose RPD was 2.2. Then, the models were used to calculate the water content and springiness for each pixel in the hyperspectral image, and a visual distribution map was generated for the visual detection of the water content and springiness of preserved eggs. Finally, the competitive adaptive weight sampling method was used to select the characteristic wavelengths, and a back propagation (BP) neural network model was established for quality prediction. The total recognition accuracy of the model was 98.3%.

In this study, hyperspectral imaging was used for nondestructive detection of preserved eggs at different maturity levels during the pickling period. First, the optimal waveband was determined based on the one-dimensional spectra and two-dimensional correlation spectra in the time-series mode, separately. Then, the modeling effects of traditional machine learning and the improved ResNet20_SE model in the optimal waveband were compared, and the results showed that the improved ResNet20_SE model was better; the overall recognition accuracy was 97.29% for the synchronous spectral dataset, and the average detection speed for a single image was 24.62 ms. Finally, the better synchronous spectral dataset ResNet20_SE model was applied to the hyperspectral pixel spectral image to calculate the value of each pixel point, and a pseudo-color technique was used for the visual detection of the spatial distribution of preserved egg maturity during the pickling process. The results of this study showed that hyperspectral imaging combined with deep learning is useful for nondestructive detection of preserved egg maturity during curing, which can lay a theoretical foundation for high-throughput online sorting of preserved egg maturity in the future.