Non-destructive detection of pork myoglobin content based on hyperspectral technology

Abstract: Hyperspectral imaging system can widely be expected to acquire a set of sample images within certain spectral bands in each pixel at the same time. In this study, rapid detection was proposed for the myoglobin content in pork samples using spectral images and deep learning. The pork was placed under the cold storage conditions at 4°C, where a total of 250 pork samples were settled at different times (0-5 d). A hyperspectral imager was used to collect the pork hyperspectral images (400 to 1 000 nm). ENVI5.3 software was also selected to determine the region of interest (ROI) in the hyperspectral images, thereby extracting the full-band average spectrum and principal component image of ROI. Subsequently, a Savitzky-Golay (SG) filter was used to denoise the spectral information for the curve smoothness and spectral resolution. A convolutional auto encoder (CAE) was utilized to extract spectral depth features. A prediction model was finally established for the content of deoxymyolglobin (DeoMb), oxymyoglobin (OxyMb), and metmyoglobin (MetMb) in the pork samples. The results showed that the determination coefficients of test datasets were 0.923 8, 0.920 3, and 0.909 2, and the root mean square errors (RMSE) were 0.033 4, 0.619 7, and 0.809 1, respectively. Furthermore, the image information of adjacent wavelengths was highly correlated against the image extraction and storage. Principal Component Analysis (PCA) was utilized to reduce the dimension of hyperspectral images for better storage and processing. As such, the images under all bands were linearly combined to form a principal component image in the ENVI5.3 software. The first three principal component images represented 90.62% of the original hyperspectral image, where the contribution rate of the first principal component was 88.50%, indicating the most information. Therefore, the first principal component image was selected for the subsequent image extraction. The first principal component image was unified to the size of 16×16 pixels, and then converted into a 768-dimensional column vector for the extraction of image depth features using a convolutional encoder. DeoM, OxyMb, and MetMb content prediction models were established using image depth features, in which the determination coefficients of test datasets were 0.772 1, 0.828 7, and 0.825 4, while the RMSE of prediction were 0.105 8, 1.302 7, and 1.566 7. The spectral and image features were fused at the data level, and then the fusion data was input into the CAE to extract the deep fusion features. The DeoMb, OxyMb, and MetMb content prediction models were also established using the fusion depth features. The determination coefficients of test datasets were 0.964 5, 0.973 2, and 0.958 5, while the RMSE of prediction were 0.015 8, 0.226 6, and 0.381 6. Obviously, the determination coefficients of the test dataset were improved, while the RMSE were reduced, compared with the individual image and spectrum information. Partial least square regression (PLSR) and support vector machine regression (SVR) prediction models were also established to further verify the relationship between the graph-spectrum fusion feature and pork myoglobin. It was found that the determination coefficients of the test dataset were greater than 0.85. Consequently, the convolutional autoencoder can be expected to extract the deep fusion features of image and spectral information. Moreover, the fusion features can better reflect the internal and external information of pork. The CNN regression model using the fusion features can also be used to improve the prediction accuracy. This finding can provide a new better way to detect the myoglobin content in pork using hyperspectral imaging.