Hyperspectral Imaging (HSI) systems are an advanced form of imaging system, which have been used in a plethora of research areas such as remote sensing [1
], food science, chemistry [3
], medical imaging [4
] and other raw materials [5
]. HSI systems eliminate the use of traditional imaging systems such as monochromic images, 3-channel images (RGB), and Multi-spectral imaging techniques in terms of accuracy, processing, and spatial information [6
]. Thus, the combination of spatial and spectral information across the electromagnetic spectrum provides a unique signature for its material, which helps in inspection, authentication, classification, and identification of different materials.
The HSI system also provides the visualization of its material, utilizing an image alongside the distribution of its chemical components. By definition, HSI provides information in the form of cubical structure, i.e., 3-dimensional cube (called Spectral cube);
represents spatial coordinates,
explains the fundamental wavelength of an image and R
represents the reflectance of minced meat. This entire electromagnetic spectrum (wavelength (
)) generates a chemical composition (nature) of the material via the intensity of the reflected light. Moreover, the HSI system forms a stack of images throughout its entire wavelength, where each image is a gray-scale (monochromic) and commonly represented as a band at each point of wavelength (
) and defined as
As earlier explained, HSI has been used for many real-world applications, and one of them is minced meat identification and classification. Minced meat is enriched with nutritional values such as iron and protein, which contains all the essential vitamins required for human growth. Minced meat is widely used in a variety of food products such as meatballs, burgers, and sausages, etc. Therefore, minced meat identification to prevent substitution is getting more attention from the research community [8
]. Accurate identification of minced meat not only minimizes the issues (risk) of human health (gastric cancer, allergies, blood pressure, and type-2 diabetes) [10
], but it also helps in investigating substitution fraud [11
]. Substitution is the most common form of fraud due to its incredible impact on economic gain, i.e., inexpensive meat will give more revenue than expensive meat, which ultimately affects consumer religion, wealth, and lifestyle [13
HSI processing used traditional Artificial Intelligence (AI) and Machine Learning (ML) methods to prevent minced meat frauds [15
]. These methods are faster and efficient [16
] than laboratory-based control measures such as pH meter, Drip-loss, and water holding capacity (WHC) [18
]. For instance, Sanz et al. [16
] introduced a study to differentiate and classify between four types of lamb muscles. The principal component analysis (PCA) was applied to reduce the dimensions of the HSI system. For the selected wavelengths, linear discrimination analysis outperformed the linear least mean squares classifier, multilayer perceptron with scaled conjugate gradient, variant of the Support Vector Machine (v-SVM), K
-Nearest Neighbor (KNN), and Logistic Regression (LR). Velas et al. [19
] proposed a method to classify the marbling of 35 beef samples. The sample types include longissimus dorsi
and are examined through the HSI sensor. For better classification, each muscle was examined through the Decision Tree (DT). Similarly, Barbon et al. [20
] used the DT approach to identify and classify chicken meat based on the REP Tree method and achieved an overall precision of
Minced meat identification process usually include color, pH and texture parameters [17
] analysis. For example, Liu et. al. [21
] used quality parameters such as; Moisture Content (MC), and the color of beef meat was examined for (0–75 s) through microwave heating treatment. Color parameters (L*, a*), and their related Maygpbin (Mb) percentage were investigated for the mean spectrum of beef samples. The prediction model—Partial Least Square Regression Model (PLSR) coefficient of determination for prediction data
= 0.890, Root Mean Square Error (RMSEP) = 0.735, and Residual Predictive Deviation (RPD) = 2.733 were established for a* values. In another work, Liu et al. [22
] investigated the quality of pork meat for predicting the color and pH of salted meat. To reduce the dimensions of the HSI cube, PCA was used and the results show L* =
, a* =
, and pH =
for the coefficients of determination (
, cv = cross-validation) and
respectively. Similarly, in 2020, Xu et al. [23
] experimented on an entire HSI cube of real salmon fillets through several feature descriptors such as Gray-level Co-occurrence Matrix (GLCM), variogram, Histograms of Oriented Gradients (HOG), and Local Binary Pattern (LBP). The obtained results showed HOG features outperformed with
accuracy for the red meat. In addition, Ayaz et al. [24
] performed the classification of multiple minced meat types (beef, mutton, and chicken). To perform better classification, myoglobin spectral features were used with the SVM algorithm which outperformed the mostly used method by achieving an overall accuracy of 88.8%.
However, in most cases, traditional approaches applied to complex 3D HSI cube require several feature extraction steps [23
], which reduces the effectiveness of the classification model [25
]. Besides, they are also limited to the use of spectral information [26
]. Irrespective of these methods, Convolutional Neural Network (CNN) allows both spatial and spectral information of HSI-system [15
]. For instance, Al-Sarayreh et al. [6
] proposed an in-depth model for beef, lamb, pork, and fat adulteration based on SVM, 1D and 3D CNN [28
] for identification and classification of different meat types. The result indicates the 3D CNN model outperforms with
, as compared to the traditional techniques. Similarly, in another work, Al-Sarayerh et al. [29
] used 3D CNN to classify red-meat using snapshot HSI.
Chunk meat classification and identification have been explored through CNN models and showed remarkable results as compared to the traditional methods. However, CNN has not yet been used for minced meat identification and classification. Therefore, this work proposes the identification of minced meat through classification using Deep Learning (DL) techniques, which follow the process of pixel-based classification of minced meat. This work also proposes the novel isos-bestic
point band reduction approach using myoglobin (Mb) pigments irrespective of traditional wavelength reduction techniques. In our previous work [24
], we have utilized the absorption feature spectrum of minced meat to perform spectral classification and used the entire spectrum rather than reduced spectra. In a nutshell, the following steps are followed.
Collection and mincing of meat types using two cross blades.
HSI system is used for image acquisition and image correction.
Reduction of dimensionality using Isos-bestic point in Mb pigments to retain maximum color information.
Classification of minced meat types using KNN, SVM, and 3D-CNN.
The rest of the paper is structured as; Section 2
, sampling process, and HSI-sensor data acquisition is explained with image correction and pre-processing. This section also explains the Isos-bestic
point wavelength reduction along with the classification techniques. Section 3
, discuss the information found in the minced meat spectrum, alongside achieved classification results for both traditional and state-of-the-art DL methods. Finally, Section 4
concludes the work with future directions.