In this study, a 400-1000nm hyperspectral camera was used to detect the inside of walnut, and FS-13, a product of Hangzhou CHNSpec Technology Co., Ltd, could be used for related research. To detect walnut surface in the spectral range of 800-1700nm, FS-15 hyperspectral camera in the spectral range of 900-1700nm can be used with wavelength resolution better than 2.5nm and up to 1200 spectral channels. The acquisition speed can reach 128FPS in the full spectrum, and the maximum after band selection is 3300Hz (support multi-region band selection).
Walnuts are a nut food suitable for all ages and an important woody oil crop. The planting area and yield of walnuts in China rank first in the world. The quality testing and grading of walnut kernels is an important link in walnut production and processing. According to relevant national standards, the appearance quality indicators of walnut kernels include integrity and skin color, while the internal quality indicators include fat content and protein content. In actual production, walnut kernel grading mainly relies on manual selection of appearance and color, which has high production costs and high randomness in grading, making it difficult to distinguish internal quality. Traditional chemical testing is destructive to samples and takes a long time to detect, making it difficult to adapt to modern production requirements. At present, research on the use of hyperspectral technology for walnut quality detection mainly focuses on the classification of walnut shells and kernels, and there have been no relevant reports on the quality of walnut kernels.
In order to explore a method to simultaneously realize the internal quality detection and appearance classification of walnut kernel, this study used hyperspectral imaging technology to screen the characteristic spectra of fat content, protein content and color of walnut kernel, and screened out the relevant characteristic bands of quality indicators in order to provide reference for the application of nondestructive testing of walnut kernel quality.
The average spectral information of walnut kernel samples in the near-infrared region (863-1704 mm) and the pre processed spectral information are shown in Figure 3. The overall characteristics of the original spectral information of the samples are basically consistent, except for the absorption peaks of water, the absorption peaks of other components are not obvious, and further processing of the spectra is needed. The preprocessing method combining MSE and SNV eliminates the influence of some background noise, making the spectral information of the sample smoother. At the same time, it further enhances the consistency of spectral information, highlights spectral peaks and valleys, and strengthens spectral features.
The appearance grade classification of walnut kernel based on spectral information and image features. Figure 6 shows the average spectral curve of three color walnut kernel samples in visible light and short wave near-infrared regions (382~1027nm). Since the noise in the front and back segments of the spectrum has a large impact, 20 waveband points in the front and back segments are removed. From Figure 6, it can be seen that in the original spectrum, the spectral reflectance of walnut kernel samples with three different colors shows a significant downward trend in the visible light range as the color changes from light to deep, and the spectrum is relatively disordered in the near-infrared range. The spectral information preprocessed by the combination of MSC and SNV methods shows certain regularity and consistency in spectral reflectance, which is helpful for subsequent spectral processing.
Using hyperspectral imaging technology, a method for detecting the internal and external quality of walnut kernels was studied. By combining spectral and image information, protein and fat content prediction of walnut kernels and appearance quality grading based on integrity and color were achieved. The results show that the combination of CARS algorithm and correlation coefficient method effectively removes irrelevant and redundant information in the full spectral band. Compared with the full spectral band, the validation set R of the feature band prediction model for protein content ² From 0.66 to 0.91, RMSEP decreased from 1.37% to 0.78%; Validation set R for fat content ² From 0.83 to 0.93, RMSEP decreased from 0.98% to 0.47%, indicating that the selected feature bands effectively reduced the complexity of the model and improved its predictive ability. By combining color difference feature spectra with image statistical feature parameters, the total color difference feature band spectra were extracted from hyperspectral images, which can significantly reduce the interference of redundant information and improve modeling efficiency. By combining the total color difference feature band spectrum with image statistical feature parameters, the classification accuracy is further improved compared to the RGB band. When using the color classification model established by the DT algorithm, the model has the highest classification accuracy (98.6%). The use of hyperspectral images simultaneously achieved the detection of internal quality parameters (protein content, fat content) and the classification of appearance quality (integrity, color) of walnut kernels, providing a new solution for the application of non-destructive testing of walnut kernel quality.
