Kiswanto*
| Hadiyanto Hadiyanto | Eko Sediyono
© 2025 The authors. This article is published by IIETA and is licensed under the CC BY 4.0 license (http://creativecommons.org/licenses/by/4.0/).
OPEN ACCESS
The purpose of this research is to evaluate the combination of Wavelet and Gray Level Co-occurrence Matrix (GLCM) methods in extracting texture features of various types of meat, including beef, buffalo, lamb, horse, and pork. This method integrates the advantages of wavelet transform in capturing spatial-frequency features with GLCM’s ability to analyse statistical texture patterns. The classification process is carried out using the k-Nearest Neighbors (k-NN) algorithm, and the model accuracy is evaluated using a confusion matrix. The research results show that the combination of Wavelet and GLCM features significantly improves the classification performance, with an average accuracy of 97.2%. Further analysis shows that the integration of these two methods provides better classification results between fresh, frozen, and rotten categories for each type of meat. Although there are some classification errors, the overall results show the reliability and effectiveness of this approach. Further research can explore parameter optimization or the integration of more sophisticated classification algorithms.
wavelet transform, gray level co-occurrence matrix (GLCM), meat texture classification, machine learning
The meat industry is an important sector in meeting food needs worldwide. Meat quality is greatly influenced by various factors, such as the type of animal, the age of the animal, the slaughtering process, and the processing and storage methods [1]. One aspect of quality that is very important and can affect consumer acceptance is meat texture. Meat texture, which includes tenderness, roughness, elasticity, and density, greatly affects consumer acceptance of the product [2]. In addition, the physical condition of the meat, whether fresh, frozen, or rotten, also plays a major role in determining the quality of the texture that can be observed in the meat image [3]. Changes in this physical condition will change the texture properties of the meat, which can be observed through digital image processing techniques [4].
Fresh meat tends to have a chewy, soft, and elastic texture, with a smooth surface and little or no wrinkles. In contrast, frozen meat has a harder, stiffer, and more porous texture due to freezing, which causes changes in muscle fibres and the structure of meat tissue. Rotten meat, on the other hand, has a softer, more watery, and disintegrating texture due to the decomposition processes that damages muscle fibres and tissue. Therefore, the analysis of meat texture is not only limited to the type of meat, but also needs to consider changes in texture that occur due to differences in physical conditions, such as fresh, frozen, or rotten.
Traditionally, meat texture quality assessment is carried out through physical methods involving laboratory testing or sensory tests conducted by experts. However, these methods require a lot of time, money, and effort. With the advancement of technology, digital image analysis has emerged as a more efficient and accurate solution for automatically assessing meat texture [5].
Research related to meat image classification has been conducted in recent years by several researchers. The ResNet152V2 algorithm was employed to classify beef, lamb and pork types, achieving 80% accuracy on a dataset of 585 images [6].
Beef quality can be checked visually by observing the colour or texture of beef using the human eye. Beef quality can be evaluated visually by examining the colour or texture of the meat using direct observation by the human eye. Although this manual method is relatively simple, it is very subjective due to differences in understanding the characteristics of fresh or damaged beef, as well as variations in the level of accuracy. The use of a combination of colour and texture features for meat image classification was carried out using Support Vector Machines, statistical approaches and the GLCM method for the feature extraction process with a total data used of 480 images, obtaining the highest accuracy of 97% [7]. Furthermore, A combination of HSV (hue, saturation, value) color features and LOOP (Local Optimal-Oriented Pattern) texture features was employed, achieving 98%-100% accuracy on a dataset of 400 images [8].
These researches have produced a high level of accuracy in identifying the quality of certain meats, namely beef and pork. However, these researches have not used variations in image data for other types of meat, namely buffalo, lamb, and horse meat. The quality studied is still limited to fresh and damaged (rotten) meat.
For this reason, this research proposes an evaluation of the combination model of Wavelet and GLCM for the extraction of texture image features of various types of meat (beef, lamb, buffalo, horse, pork) based on quality, namely fresh, frozen and rotten.
This research contributes to improving the accuracy of meat image texture classification based on its quality. By using a combination of Wavelet and GLCM, this research can provide more precise results in identifying meat quality, whether it is fresh, frozen, or rotten. This is important for monitoring the quality of meat products, which in turn can be applied in automation systems for inspection and quality control in the food industry, thereby helping manufacturers ensure more consistent product quality.
In addition, the combination of these two methods is expected to produce a more accurate and robust texture classification system. Wavelet transform will capture texture patterns globally and multi-scale, while GLCM applied to the results of wavelet decomposition will extract more specific local spatial relationships.
Wavelet transform has been used in several research for texture feature extraction from images. Wavelet allows image decomposition into various frequencies and scales, which is useful in identifying more complex texture patterns in images [9]. Wavelet transform has been applied to identify young green tea leaves in color images [10]. For improved material and texture classification in object identification, wavelet transform has been combined with Convolutional Neural Networks (CNNs) [11]. In document image classification, a novel approach integrating wavelet transform with Swin Transformer has been proposed [12].
The measurement parameters of beef tenderness cannot be used for visual assessment, because they do not have visual standards. In the study [13], a method was developed to extract beef texture as a basic feature in the beef classification process using statistical analysis of GLCM and frequency domain of Discrete Wavelet Transform (DWT). The findings of this research indicate that beef texture, measured using a feature extraction method based on GLCM and DWT, can be used to assess beef tenderness effectively.
The future work in this research focuses on evaluating various texture features produced by both methods, including contrast, correlation, energy, and homogeneity from GLCM, as well as horizontal, vertical, and diagonal details from wavelet. Unlike previous research that was limited to only one type of meat or type of extraction method, this research expands the scope by testing the combination of Wavelet and GLCM techniques for various types of meat and assessing the model’s ability to provide more accurate and efficient results in meat quality classification.
Empirical evidence from previous studies shows that the use of texture features from GLCM individually is indeed able to provide strong local statistical information. However, when GLCM is combined with multi-resolution information from wavelet transform, the classification performance can be significantly improved. In addition, wavelet transform has advantages in analyzing non-stationary signals and complex texture structures because of its ability to extract information at various scales and frequencies simultaneously [14]. This makes wavelet very suitable for image processing applications that require the detection of fine patterns and texture changes, such as in meat images. By combining the power of GLCM spatial analysis and the multi-resolution capabilities of wavelet, this approach is believed to be able to produce a more adaptive and accurate texture classification system for various types of meat.
This research was conducted systematically for the classification of meat image textures based on their quality by utilizing digital image processing technology. Meat images were collected and grouped into three types of meat (beef, lamb, buffalo, horse, pork) with the quality categories studied, namely fresh, frozen and rotten.
Figure 1 shows a block diagram of the research stages that will be carried out.
Figure 1. Research stages
3.1 Data collection
The meat type dataset in this research was taken using a digital camera, photographed with the same light intensity and distance. There are 5 types of meat, namely beef, buffalo, lamb, horse and pork. Each type of meat was taken as many as 50 meat images and then photographed one by one, so that the total data used was 750 images.
Table 1 shows sample images of various types of meat with fresh, frozen, and rotten qualities.
Based on Table 1, it can be observed that each image composition of meat types (beef, buffalo, lamb, horse, and pork) has the same number of samples for each category (fresh, frozen, and rotten). Each category has 50 samples per type of meat, so the total samples per type of meat is 150 samples. So the total samples of all types of meat are: 150 samples/meat ×5 categories of meat types =750 samples.
Table 1. Image samples of different types of meat (fresh, frozen, and rotten)
|
Meat Types |
Category |
Image |
Number of Samples |
|
Beef |
Fresh |
50 |
|
|
Frozen |
50 |
||
|
Rotten |
50 |
||
|
Buffalo |
Fresh |
50 |
|
|
Frozen |
50 |
||
|
Rotten |
50 |
||
|
Lamb |
Fresh |
50 |
|
|
Frozen |
50 |
||
|
Rotten |
50 |
||
|
Horse |
Fresh |
50 |
|
|
Frozen |
50 |
||
|
Rotten |
50 |
||
|
Fork |
Fresh |
50 |
|
|
Frozen |
50 |
||
|
Rotten |
50 |
||
|
Total Samples: |
750 |
||
3.2 Pre-processing
The first stage carried out in image pre-processing is cropping. This process is the cutting of pixels in a digital image according to the specified position and dimensions [14]. This cutting process is carried out based on spatial coordinates in vector format [x, y, width, height], which includes all pixels within the specified area. Furthermore, resizing is carried out to change the image size from 6000 × 4000 pixels to 100 × 500 pixels. This step aims to reduce computing time and the load on the system. The last stage is the conversion of the colour image to grayscale. This process is important for feature extraction using the GLCM to ensure more accurate texture analysis. Sample images produced from pre-processing can be seen in Figure 2.
Figure 2. Pre-processing: (a). Original beef image sample, (b). Cropped image, (c). Resized image 100 × 500, (d). Convert color image to grayscale image
3.3 Feature extraction
Feature extraction is the process of extracting features from an image that are needed for classification purposes [15]. In this research, feature extraction was carried out using two main methods, namely wavelet transformation (including Haar wavelet, Daubechies, Coiflets, Symlet) and GLCM.
In wavelet transform, the image is decomposed into several coefficients that describe the intensity changes at various scales and orientations. These coefficients include LL (Low-Low), LH (Low-High), HL (High-Low), and HH (High-High). The LL coefficients store the average or approximation information of the image at low scales, describing the basic structure of the image. Meanwhile, the LH, HL, and HH coefficients store the image details in the horizontal, vertical, and diagonal directions with different resolutions, each describing the texture changes in the image. By using various types of wavelets such as Haar, Daubechies, Coiflet, and Symlet, these coefficients can provide different descriptions of the image texture, depending on the wavelet’s ability to capture details at various frequency levels [16, 17].
Furthermore, GLCM functions to calculate the frequency of occurrence of pixels with a certain gray intensity that are adjacent to other pixels in various directions, including horizontally, vertically, and diagonally right and left [18]. The feature extraction process produces statistical parameters, namely contrast, correlation, energy, homogeneity and entropy [19, 20]. Energy can be calculated using Eq. (1).
$Energy=\sum_i \sum_j P(i, j)^2$ (1)
Contrast, also known as inertia, is a measure that describes the degree of variation in the intensity of different pixels in an image. Contrast can be calculated by Eq. (2).
$Contrast =\sum_i \sum_j(i-j)^2 * P(i, j)$ (2)
Homogeneity is a measure that reflects the level of matrix elements that are close to the diagonal of the GLCM matrix. The homogeneity value will be high if the matrix elements that have high values are located near the diagonal. Homogeneity can be calculated using Eq. (3).
$Homogeneity=\sum_i \sum_j \frac{P(i, j)}{1+(i-j)^2}$ (3)
Entropy is a measure of the complexity of the texture or information contained in an image. Entropy can be calculated using Eq. (4).
$Entropy=-\sum_i \sum_j P(i, j) * \log P(i, j)$ (4)
Correlation is a measure that describes the degree of relationship between two pixels. Correlation can be calculated using Eq. (5).
Corelation $=\frac{\sum_i \sum_j * P(i, j)-\mu_{x *} \mu_y}{\sigma_x * \sigma_y}$ (5)
3.4 Feature fusion
Feature Fusion is a technique in machine learning and computer vision that combines features from multiple sources or multiple levels of representation to improve model performance [21]. This process is often used in a variety of applications, such as image classification, object detection, and segmentation, where more and more diverse information can provide a better understanding of the data [22].
3.5 Classification with k-NN
Classification in machine learning is one of the tasks that aims to map input data into certain categories or classes based on patterns in the data [23]. One of the methods commonly used in classification is k-NN. k-NN is an instance-based learning algorithm used for classification and regression. In the case of classification, this algorithm groups data into certain classes based on their proximity to other data that has been labeled [24, 25].
3.6 Evaluation
Evaluation in machine learning is a very important step to assess the performance of a trained model and measure how well the model is able to solve the problem at hand [26]. Evaluation aims to find out whether the model can generalize well on data that has never been seen before and whether the model is robust enough to handle various conditions.
In the evaluation process, the data must first be divided into several subsets, namely training data to train the model, validation data to select the best model, and test data used to test the model’s performance after training [27, 28]. Next, evaluation metrics are used to assess the model’s results. In image classification, metrics such as accuracy, precision, recall, and F1-score are often used. Accuracy measures how often a model makes correct predictions, while precision and recall assess the model’s ability to identify the correct class.
This section presents the research results and analysis on the evaluation of the combination model of Wavelet and GLCM for the extraction of texture image features of various types of meat.
To conduct testing, a set of collected datasets was divided into 80% training data and 20% test data. This testing was conducted by comparing the results of feature extraction and meat classification using a single method (only wavelet or only GLCM) and a combination of both.
4.1 Wavelet
The test scenario is compiled based on various types of meat (beef, buffalo, lamb, horse, fork) which each have different condition categories (fresh, frozen, rotten), then tested using four different types of wavelets, namely Haar (H), Daubechies (D), Coiflets (C), and Symlet (S). This scenario aims to evaluate the performance of the model in extracting texture features of meat images in various conditions and types of meat using the wavelet transform technique. Table 2 shows the results of feature extraction using wavelet transform
Based on the results of feature extraction using wavelet transform presented in Table 2, it can be seen that the wavelet transform technique using four types of wavelets, namely Haar, Daubechies, Coiflets, and Symlet, has different performance in extracting texture features of meat images in various types and conditions. In general, Haar Wavelet gives the highest average results (87.42%), followed by Daubechies (85.81%), Coiflets (83.32%), and Symlet (83.26%). This indicates that Haar is superior in capturing texture information compared to other wavelets because Haar has a high ability to recognize small differences in image surface patterns, which are often important indicators in distinguishing meat conditions. Meanwhile, the performance of Coiflets and Symlets is almost the same, but still lower than Haar and Daubechies.
Table 2. Feature extraction results using wavelet transform
|
Type of Meat |
Category |
Wavelet |
|||
|
H |
D |
C |
S |
||
|
Beef |
Fresh |
89.21 |
87.12 |
85.46 |
85.65 |
|
Frozen |
88.76 |
86.89 |
84.12 |
83.26 |
|
|
Rotten |
96.82 |
85.44 |
82.76 |
82.21 |
|
|
Buffalo |
Fresh |
89.56 |
85.23 |
83.29 |
83.21 |
|
Frozen |
86.33 |
83.56 |
81.19 |
81.86 |
|
|
Rotten |
87.67 |
82.11 |
79.28 |
79.01 |
|
|
Lamb |
Fresh |
84.78 |
82.43 |
80.27 |
80.14 |
|
Frozen |
82.11 |
80.66 |
78.73 |
78.34 |
|
|
Rotten |
80.97 |
78.32 |
76.02 |
76.10 |
|
|
Horse |
Fresh |
91.45 |
90.43 |
88.91 |
88.28 |
|
Frozen |
89.60 |
88.89 |
86.21 |
86.80 |
|
|
Rotten |
87.24 |
86.91 |
84.54 |
84.02 |
|
|
Fork |
Fresh |
93.46 |
86.32 |
89.64 |
86.21 |
|
Frozen |
80.59 |
84.27 |
82.31 |
82.11 |
|
|
Rotten |
83.33 |
82.63 |
80.04 |
80.70 |
|
|
Average: |
87.42 |
85.81 |
83.32 |
83.26 |
|
Meanwhile, the performance of Coiflets and Symlet is almost the same, but still lower than Haar and Daubechies.
Fresh meat condition produces the highest extraction value compared to frozen and rotten conditions. This shows that the texture structure of fresh meat is easier to recognize than the texture of meat that has been frozen or rotten. In addition, the test results also show that certain types of meat, such as horse meat and pork, have a texture that is easier to identify with high extraction results, especially in the fresh category. In contrast, lamb meat showed the lowest extraction results in all categories, indicating a greater challenge in recognizing its texture.
4.2 Texture analysis of meat images using GLCM
Testing with GLCM was conducted to analyse the texture of meat images based on various types of meat, namely beef, buffalo, lamb, horse, and pork. Each type of meat was tested in three different condition categories: fresh, frozen, and rotten. This approach aims to evaluate the ability of GLCM to extract meat texture features based on a co-occurrence matrix that considers the spatial relationship between pixels in the image. In this scenario, the main parameters calculated using GLCM include several texture features, namely, Energy (EY), Contrast (CO), Correlation (CR), Homogeneity (H), and Entropy (EN)
Meat images were taken from each type and category, then processed to produce GLCM features at certain angles (0°, 45°, 90°, 135°) with certain pixel distances. The results of this feature extraction are used to compare how the texture of each type of meat is affected by different conditions. This test is expected to provide insight into the unique texture patterns of each type of meat, so that it can be used for classification purposes. Table 3 below shows the results of feature extraction using GLCM.
From the feature extraction results presented in Table 3 using GLCM at an angle of 0°, an analysis of five texture features can be carried out, namely Energy (EY), Contrast (CO), Correlation (CR), Homogeneity (H), and Entropy (EN). In general, the Energy value ranges from 0.6954 to 0.8021 with an average of 0.7463, indicating that meat with a more homogeneous texture has a higher value, such as Pork, fresh meat which has the highest value (0.8021). Conversely, Lamb Rotten meat has the lowest Energy value (0.6789), indicating a more varied texture. For contrast, the value which measures texture irregularity ranges from 10.65 to 19.11 with an average of 14.683. Fresh pork had the lowest Contrast value (10.65), indicating a more homogeneous texture, while Lamb Rotten had the highest Contrast value (19.11), and indicating greater irregularity.
The Correlation feature measures the relationship between adjacent pixels, values ranged from 0.8574 to 0.9345, with an average of 0.9028. Fresh pork had the highest Correlation value (0.9345), indicating a close relationship between pixels, while Lamb Rotten had the lowest Correlation value (0.8574), indicating a weaker relationship between pixels. Homogeneity measures the uniformity of pixel distribution, values ranged from 0.8045 to 0.8921, with an average of 0.8490. Fresh pork had the highest Homogeneity value (0.8921), indicating a very uniform distribution, while Lamb Rotten had the lowest Homogeneity value (0.7989), indicating irregularity in pixel distribution. Finally, on the Entropy feature, which describes the irregularity or complexity of the image, the values ranged from 1.1987 to 1.5342, with an average of 1.3599. Fresh pork had the lowest Entropy value (1.1987), indicating less irregularity, while Lamb Rotten had the highest Entropy value (1.5342), indicating greater variation in texture.
Table 4 shows the results of feature extraction using GLCM for the 45° angle.
Based on the feature extraction results that use GLCM at a 45° angle in Table 4, the analysis of the five texture features showed significant differences in texture characteristics between types and conditions of meat. Energy which indicates texture homogeneity, had the highest value in Fresh pork (0.8045), indicating a very homogeneous texture, while Lamb Rotten had the lowest value (0.6833), indicating a more varied texture. Contrast, which measures texture irregularity, also showed a striking difference, with Fresh pork having the lowest value (10.20), indicating a more uniform texture, and Lamb Rotten having the highest value (18.29), indicating greater texture irregularity.
In the Correlation feature, which measures the relationship between adjacent pixels, Fresh pork shows a stronger relationship between pixels with the highest value (0.9358), while Lamb Rotten has the lowest value (0.8581), indicating a weaker relationship between pixels. Homogeneity, which measures the uniformity of pixel distribution, shows that Fresh pork has a very uniform distribution with the highest value (0.8905), while Lamb Rotten has the lowest value (0.7968), indicating a more irregular pixel distribution.
Table 3. Feature extraction results using GLCM (Angle 0°)
|
Type of Meat |
Category |
GLCM |
||||
|
EY |
CO |
CR |
H |
EN |
||
|
Beef |
Fresh |
0.7823 |
12.34 |
0.9231 |
0.8745 |
1.2075 |
|
Frozen |
0.7564 |
14.67 |
0.9045 |
0.8423 |
1.3161 |
|
|
Rotten |
0.7121 |
16.89 |
0.8867 |
0.8156 |
1.4547 |
|
|
Buffalo |
Fresh |
0.7645 |
13.21 |
0.9178 |
0.8632 |
1.2894 |
|
Frozen |
0.7342 |
15.45 |
0.8943 |
0.8327 |
1.3995 |
|
|
Rotten |
0.6954 |
18.02 |
0.8712 |
0.8045 |
1.5123 |
|
|
Lamb |
Fresh |
0.7431 |
14.12 |
0.9102 |
0.8523 |
1.3102 |
|
Frozen |
0.7156 |
16.34 |
0.8825 |
0.8257 |
1.4231 |
|
|
Rotten |
0.6789 |
19.11 |
0.8574 |
0.7989 |
1.5342 |
|
|
Horse |
Fresh |
0.7954 |
11.89 |
0.9289 |
0.8834 |
1.2103 |
|
Frozen |
0.7682 |
13.76 |
0.9134 |
0.8556 |
1.3294 |
|
|
Rotten |
0.7256 |
15.98 |
0.8912 |
0.8264 |
1.4405 |
|
|
Fork |
Fresh |
0.8021 |
10.65 |
0.9345 |
0.8921 |
1.1987 |
|
Frozen |
0.7794 |
12.89 |
0.9210 |
0.8654 |
1.3156 |
|
|
Rotten |
0.7412 |
14.92 |
0.9048 |
0.8423 |
1.4234 |
|
|
Average |
0.7463 |
14.683 |
0.9028 |
0.8490 |
1.3599 |
|
Table 4. Feature extraction results using GLCM (Angle 45°)
|
Type of Meat |
Category |
GLCM |
||||
|
EY |
CO |
CR |
H |
EN |
||
|
Beef |
Fresh |
0.7982 |
13.56 |
0.9262 |
0.8701 |
1.2204 |
|
Frozen |
0.7634 |
15.21 |
0.9056 |
0.8410 |
1.3407 |
|
|
Rotten |
0.7186 |
17.32 |
0.8854 |
0.8135 |
1.4587 |
|
|
Buffalo |
Fresh |
0.7695 |
12.98 |
0.9193 |
0.8617 |
1.2803 |
|
Frozen |
0.7391 |
14.99 |
0.8951 |
0.8304 |
1.3980 |
|
|
Rotten |
0.7018 |
17.43 |
0.8718 |
0.8036 |
1.5111 |
|
|
Lamb |
Fresh |
0.7480 |
13.89 |
0.9125 |
0.8507 |
1.3045 |
|
Frozen |
0.7204 |
15.87 |
0.8832 |
0.8232 |
1.4205 |
|
|
Rotten |
0.6833 |
18.29 |
0.8581 |
0.7968 |
1.5330 |
|
|
Horse |
Fresh |
0.7991 |
11.32 |
0.9300 |
0.8852 |
1.2058 |
|
Frozen |
0.7719 |
13.43 |
0.9142 |
0.8541 |
1.3241 |
|
|
Rotten |
0.7283 |
15.57 |
0.8927 |
0.8275 |
1.4402 |
|
|
Fork |
Fresh |
0.8045 |
10.20 |
0.9358 |
0.8905 |
1.1950 |
|
Frozen |
0.7810 |
12.32 |
0.9215 |
0.8643 |
1.3140 |
|
|
Rotten |
0.7440 |
14.58 |
0.9054 |
0.8419 |
1.4231 |
|
|
Average |
0.7484 |
14.79 |
0.9027 |
0.8495 |
1.3637 |
|
Finally, Entropy, which describes the complexity or irregularity of texture, shows that Fresh pork has the lowest value (1.1950), indicating a more regular texture, while Lamb Rotten has the highest value (1.5330), indicating a more complex variation in texture. Overall, Fresh pork shows a more homogeneous and regular texture, while Lamb Rotten shows a more irregular and complex texture.
Table 5 shows the feature extraction results for the 90° angle.
Table 5. Feature extraction results using GLCM (Angle 90°)
|
Type of Meat |
Category |
GLCM |
||||
|
EY |
CO |
CR |
H |
EN |
||
|
Beef |
Fresh |
0.7834 |
13.45 |
0.9241 |
0.8752 |
1.2403 |
|
Frozen |
0.7582 |
14.60 |
0.9051 |
0.8431 |
1.3520 |
|
|
Rotten |
0.7153 |
16.72 |
0.8860 |
0.8175 |
1.4635 |
|
|
Buffalo |
Fresh |
0.7672 |
13.05 |
0.9185 |
0.8612 |
1.2815 |
|
Frozen |
0.7375 |
15.34 |
0.8938 |
0.8321 |
1.3925 |
|
|
Rotten |
0.6984 |
17.49 |
0.8723 |
0.8042 |
1.5140 |
|
|
Lamb |
Fresh |
0.7461 |
14.00 |
0.9115 |
0.8521 |
1.3084 |
|
Frozen |
0.7192 |
16.15 |
0.8839 |
0.8243 |
1.4237 |
|
|
Rotten |
0.6819 |
18.42 |
0.8564 |
0.7983 |
1.5345 |
|
|
Horse |
Fresh |
0.7988 |
11.80 |
0.9292 |
0.8845 |
1.2101 |
|
Frozen |
0.7710 |
13.60 |
0.9133 |
0.8560 |
1.3220 |
|
|
Rotten |
0.7251 |
15.79 |
0.8917 |
0.8287 |
1.4415 |
|
|
Fork |
Fresh |
0.8032 |
10.90 |
0.9353 |
0.8912 |
1.1973 |
|
Frozen |
0.7801 |
12.45 |
0.9208 |
0.8665 |
1.3158 |
|
|
Rotten |
0.7403 |
14.75 |
0.9041 |
0.8412 |
1.4239 |
|
|
Average |
0.7454 |
14.73 |
0.9025 |
0.8442 |
1.3524 |
|
Table 6. Feature extraction results using GLCM (Angle 135°)
|
Type of Meat |
Category |
GLCM |
||||
|
EY |
CO |
CR |
H |
EN |
||
|
Beef |
Fresh |
0.7834 |
12.12 |
0.9235 |
0.8750 |
1.2632 |
|
Frozen |
0.7560 |
14.25 |
0.9045 |
0.8400 |
1.3780 |
|
|
Rotten |
0.7125 |
16.85 |
0.8862 |
0.8130 |
1.4578 |
|
|
Buffalo |
Fresh |
0.7643 |
13.05 |
0.9180 |
0.8600 |
1.2855 |
|
Frozen |
0.7321 |
15.30 |
0.8939 |
0.8300 |
1.3960 |
|
|
Rotten |
0.6950 |
17.20 |
0.8700 |
0.8020 |
1.5125 |
|
|
Lamb |
Fresh |
0.7425 |
13.45 |
0.9115 |
0.8510 |
1.3120 |
|
Frozen |
0.7150 |
15.50 |
0.8830 |
0.8240 |
1.4210 |
|
|
Rotten |
0.6780 |
18.10 |
0.8560 |
0.7980 |
1.5330 |
|
|
Horse |
Fresh |
0.7948 |
11.78 |
0.9290 |
0.8840 |
1.2080 |
|
Frozen |
0.7685 |
13.60 |
0.9130 |
0.8550 |
1.3280 |
|
|
Rotten |
0.7250 |
15.70 |
0.8915 |
0.8270 |
1.4410 |
|
|
Fork |
Fresh |
0.8015 |
10.15 |
0.9355 |
0.8915 |
1.1990 |
|
Frozen |
0.7790 |
12.50 |
0.9215 |
0.8640 |
1.3155 |
|
|
Rotten |
0.7410 |
14.80 |
0.9050 |
0.8420 |
1.4235 |
|
|
Average |
0.7465 |
14.27 |
0.9027 |
0.8492 |
1.3655 |
|
Based on the results of feature extraction using GLCM for a 90° angle in Table 5, it can be analysed that the condition of the meat affects each extracted feature. For the Energy (EY) feature, the average value is 0.7454, which indicates the level of regularity of the image texture. Fresh meat tends to have a higher Energy value, indicating a more consistent texture, while rotting meat shows a lower Energy value due to its more varied texture. The Contrast (CO) feature, with an average of 14.73, measures the difference in pixel intensity, where rotting meat shows a higher Contrast value, reflecting a rougher and more non-uniform texture. Conversely, fresh and frozen meat has lower Contrast values, indicating a smoother texture.
For Correlation (CR), which averages 0.9025, fresh meat shows stronger and more regular pixel-to-pixel relationships, with the highest value in Fresh Beef. Rotting meat, with a more irregular texture, has a lower Correlation value. Homogeneity (H), with an average value of 0.8442, describes the uniformity of the image texture. Fresh meat generally has a higher Homogeneity value, indicating a smoother and more uniform texture, while rotting meat has a lower Homogeneity value. Finally, the Entropy (EN) feature, with an average of 1.3524, measures the irregularity in the image texture. Rotting meat shows a higher Entropy value, reflecting a more complex and chaotic texture, while fresh meat has lower Entropy, indicating a more regular texture.
Table 6 shows the feature extraction result for 135° angle.
Based on the feature extraction results using GLCM for an angle of 135° in Table 6, there are several patterns that can be identified related to the condition and type of meat. For the Energy (EY) feature, the average value of 0.7465 indicates that fresh meat tends to have a higher Energy value, reflecting a more regular and consistent texture, while rotting meat shows a decrease in Energy value, indicating a more irregular and more complex texture. Contrast (CO), with an average of 14.27, shows the difference in intensity between pixels. Higher Contrast values are found in rotting meat, meaning the texture becomes rougher and more contrasting, while fresh meat has a lower Contrast value with a smoother texture.
In Correlation (CR), with an average of 0.9027, fresh meat shows a stronger relationship between pixels with higher values, indicating a more regular and uniform texture. Rotting meat has a lower Correlation value, indicating a more random and unorganized texture. Homogeneity (H), with an average value of 0.8492, measures the uniformity of the image texture. Fresh meat tends to have higher Homogeneity values, while rotting meat has lower Homogeneity values, indicating irregularity in its texture structure. Finally, Entropy (EN), with an average value of 1.3655, measures the degree of irregularity in the image texture. Rotting meat generally has higher Entropy values, indicating a more complex and chaotic texture, while fresh meat shows lower Entropy, reflecting a more regular texture.
Overall, the results of GLCM feature extraction at 135° angle show that meat condition affects the texture properties of the image that can be measured through parameters such as Energy, Contrast, Correlation, Homogeneity, and Entropy. Fresh meat tends to have more consistent values in these parameters, while rotting meat shows higher values in Contrast and Entropy, and lower in Energy and Correlation, indicating a more irregular and rougher texture. This indicates that the GLCM technique at 135° angle can be effectively used to distinguish the quality and condition of meat based on image texture.
4.3 Feature fusion
In the combination of Wavelet and GLCM, features from both methods are combined to obtain a more complete description of image texture. Wavelet provides frequency information at various scales, while GLCM describes the spatial relationship between pixels based on intensity values. The combination of both can identify more texture details in the image and improve the model’s ability to distinguish image classes in the analysis of fresh, frozen, and rotten meat quality. In this context, the merging process is carried out using early fusion. Table 7 shows the results of combining the two Wavelet and GLCM methods.
Table 7. Results of combining the feature extraction methods (Wavelet and GLCM)
|
Type of Meat |
Category |
Wavelet Coefficients |
GLCM |
Combined Features |
||||||
|
Horizontal |
Vertical |
Diagonal |
EY |
CT |
CO |
H |
EN |
|||
|
Beef |
Fresh |
0.7823 |
0.7564 |
0.7121 |
0.7868 |
12.86 |
0.9242 |
0.8737 |
1.2140 |
0.7823, 0.7564, 0.7121, 0.7868, 12.86, 0.9242, 0.8737, 1.2140 |
|
Frozen |
0.7564 |
0.7342 |
0.6954 |
0.7585 |
14.68 |
0.9049 |
0.8416 |
1.3467 |
0.7564, 0.7342, 0.6954, 0.7585, 14.68, 0.9049, 0.8416, 1.3467 |
|
|
Rotten |
0.7121 |
0.6954 |
0.6789 |
0.7155 |
16.94 |
0.8861 |
0.8149 |
1.4587 |
0.7121, 0.6954, 0.6789, 0.7155, 16.94, 0.8861, 0.8149, 1.4587 |
|
|
Buffalo |
Fresh |
0.7645 |
0.7342 |
0.6954 |
0.7664 |
13.07 |
0.9184 |
0.8615 |
1.2842 |
0.7645, 0.7342, 0.6954, 0.7664, 13.07, 0.9184, 0.8615, 1.2842 |
|
Frozen |
0.7342 |
0.7156 |
0.6789 |
0.7357 |
15.27 |
0.8943 |
0.8313 |
1.3965 |
0.7342, 0.7156, 0.6789, 0.7357, 15.27, 0.8943, 0.8313, 1.3965 |
|
|
Rotten |
0.6954 |
0.6789 |
0.6256 |
0.6977 |
17.54 |
0.8713 |
0.8036 |
1.5125 |
0.6954, 0.6789, 0.6256, 0.6977, 17.54, 0.8713, 0.8036, 1.5125, |
|
|
Lamb |
Fresh |
0.7431 |
0.7156 |
0.6780 |
0.7449 |
13.87 |
0.9114 |
0.8515 |
1.3088 |
0.7431, 0.7156, 0.6780, 0.7449, 13.87, 0.9114, 0.8515, 1.3088 |
|
Frozen |
0.7156 |
0.6789 |
0.6534 |
0.7176 |
15.97 |
0.8832 |
0.8243 |
1.4221 |
0.7156, 0.6789, 0.6534, 0.7176, 15.97, 0.8832, 0.8243, 1.4221 |
|
|
Rotten |
0.6789 |
0.6534 |
0.6123 |
0.6805 |
18.48 |
0.8570 |
0.7980 |
1.5337 |
0.6789, 0.6534, 0.6123, 0.6805, 18.48, 0.8570, 0.7980, 1.5337 |
|
|
Horse |
Fresh |
0.7954 |
0.7682 |
0.7256 |
0.7970 |
11.70 |
0.9293 |
0.8843 |
1.2086 |
0.7954, 0.7682, 0.7256, 0.7970, 11.70, 0.9293, 0.8843, 1.2086 |
|
Frozen |
0.7682 |
0.7256 |
0.7156 |
0.7699 |
13.60 |
0.9135 |
0.8552 |
1.3259 |
0.7682, 0.7256, 0.7156, 0.7699, 13.60, 0.9135, 0.8552, 1.3259 |
|
|
Rotten |
0.7256 |
0.7186 |
0.6534 |
0.7260 |
15.76 |
0.8918 |
0.8274 |
1.4408 |
0.7256, 0.7186, 0.6534, |
|
|
Fork |
Fresh |
0.8021 |
0.7794 |
0.7412 |
0.8028 |
10.48 |
0.9353 |
0.8913 |
1.1975 |
0.8021, 0.7794, 0.7412, |
|
Frozen |
0.7794 |
0.7412 |
0.7156 |
0.7799 |
12.54 |
0.9212 |
0.8651 |
1.3152 |
0.7794, 0.7412, 0.7156, 0.7260, 15.76, 0.8918, 0.8274, 1.4408 |
|
|
Rotten |
0.7412 |
0.7186 |
0.6776 |
0.7416 |
14.76 |
0.9048 |
0.8419 |
1.4235 |
0.7412, 0.7186, 0.6776, 0.7416, 14.76, 0.9048, 0.8419, 1.4235 |
|
Based on Table 7, the results of combining the Wavelet and GLCM feature extraction methods produce a more complete data representation, allowing for more accurate separation of meat categories. The increasing entropy value in the rotten category indicates significant texture degradation, while the higher energy value in fresh meat reflects a more regular texture structure. This shows that feature fusion method is effective in analyzing texture images for the classification of meat types and conditions.
4.4 Classification with k-NN
The classification process using the k-NN algorithm begins by utilizing the combined feature data of the Wavelet and GLCM methods. This data includes wavelet coefficients (horizontal, vertical, diagonal) and texture features from GLCM (energy, contrast, correlation, homogeneity, and entropy). The first step is to standardize the data to ensure the same scale. Next, the dataset is divided into training data and test data for performance evaluation. The k-NN model works by calculating the distance between the test data and the training data using a distance metric such as Euclidean. The value of k=3 is used, where the three nearest neighbors determine the class based on the majority. The classification results are evaluated using accuracy, precision, recall, and confusion matrix to measure model performance. This process aims to identify meat categories (fresh, frozen, rotten) with high accuracy based on the extracted features. Table 8 shows the results of meat texture classification using the k-NN method.
Based on Table 8, the results of meat texture classification using the k-NN method showed very good performance with an average overall accuracy of 97.2%. Classification was carried out on five types of meat, namely Beef, Buffalo, Lamb, Horse, and Pork, each of which is divided into three condition categories: Fresh, Frozen, and Rotten. The results show that Beef has a high level of accuracy, ranging from 98%-100%, with the model being able to distinguish the three categories consistently. Buffalo and Lamb meat showed an average accuracy of 96%-98%, although there were small errors in distinguishing the Frozen and Rotten categories. Horse meat had the highest level of accuracy, reaching 100% in almost all categories, indicating the model’s excellent ability to classify this type of meat. Meanwhile, Pork meat also showed high performance with an accuracy of 100% for the Fresh and Frozen categories, and a slight decrease in the Rotten category, with an accuracy ranging from 94%-98%.
Table 8. Results of meat texture classification using the k-NN method
|
Sample |
Type of Meat |
Accuracy (%) |
||||||||||||||
|
Beef |
Buffalo |
Lamb |
Horse |
Fork |
||||||||||||
|
Class |
Class |
Class |
Class |
Class |
||||||||||||
|
Fresh |
Frozen |
Rotten |
Fresh |
Frozen |
Rotten |
Fresh |
Frozen |
Rotten |
Fresh |
Frozen |
Rotten |
Fresh |
Frozen |
Rotten |
|
|
|
50 |
0 |
0 |
0 |
0 |
0 |
1 |
0 |
0 |
49 |
0 |
0 |
0 |
0 |
0 |
0 |
98 |
|
50 |
0 |
0 |
0 |
0 |
0 |
0 |
1 |
49 |
0 |
0 |
0 |
0 |
0 |
0 |
0 |
98 |
|
50 |
0 |
49 |
0 |
0 |
0 |
0 |
0 |
0 |
0 |
0 |
0 |
0 |
0 |
1 |
0 |
98 |
|
50 |
0 |
0 |
0 |
0 |
0 |
2 |
0 |
0 |
0 |
0 |
0 |
0 |
48 |
0 |
0 |
96 |
|
50 |
0 |
0 |
0 |
0 |
0 |
0 |
0 |
0 |
0 |
0 |
50 |
0 |
0 |
0 |
0 |
100 |
|
50 |
0 |
0 |
48 |
0 |
0 |
2 |
0 |
0 |
0 |
0 |
0 |
0 |
0 |
0 |
0 |
96 |
|
50 |
0 |
0 |
0 |
0 |
0 |
0 |
0 |
0 |
0 |
0 |
0 |
50 |
0 |
0 |
0 |
100 |
|
50 |
50 |
0 |
0 |
0 |
0 |
0 |
0 |
0 |
0 |
0 |
0 |
0 |
0 |
0 |
0 |
100 |
|
50 |
0 |
0 |
0 |
0 |
1 |
0 |
0 |
0 |
0 |
0 |
0 |
0 |
0 |
49 |
0 |
98 |
|
50 |
0 |
0 |
0 |
0 |
0 |
0 |
2 |
0 |
0 |
48 |
0 |
0 |
0 |
0 |
0 |
96 |
|
50 |
0 |
0 |
0 |
0 |
0 |
0 |
50 |
0 |
0 |
0 |
0 |
0 |
0 |
0 |
0 |
100 |
|
50 |
0 |
1 |
0 |
49 |
0 |
0 |
0 |
0 |
0 |
0 |
0 |
0 |
0 |
0 |
0 |
98 |
|
50 |
0 |
0 |
0 |
0 |
0 |
0 |
0 |
0 |
0 |
0 |
0 |
0 |
0 |
0 |
50 |
98 |
|
50 |
0 |
0 |
0 |
0 |
0 |
50 |
0 |
0 |
0 |
0 |
0 |
0 |
|
0 |
0 |
100 |
|
50 |
0 |
3 |
0 |
0 |
47 |
0 |
0 |
0 |
0 |
0 |
0 |
0 |
0 |
0 |
0 |
94 |
|
Average |
97.2 |
|||||||||||||||
Although the overall performance is very good, small errors in distinguishing Frozen and Rotten categories, especially in Buffalo and Lamb meat, indicate the similarity in texture between these categories. Potential improvements can be made by adding training data, adjusting the k parameters of k-NN, or using additional features to improve the separability between categories. In general, the k-NN method proved effective in classifying meat texture based on features extracted from Wavelet and GLCM, providing accurate and consistent prediction results.
In the context of this research, the main reason for distinguishing frozen and rotten meat is not only important from the technical perspective of image classification, but also has major implications for health, legal compliance, industrial efficiency, and public trust.
4.5 Evaluation
Confusion matrix is an important evaluation tool in assessing the performance of classification models [29-31]. One of them is the task of classifying meat texture based on fresh, frozen, and rotten categories as shown in Figure 3.
Figure 3. Visualization of confusion matrix
The evaluation process begins with the collection of meat texture data that has been extracted using the Wavelet and GLCM methods. The resulting features are used as input for the k-NN model. The dataset is divided into training data to develop the model and testing data to evaluate the predictions. The k-NN model then predicts the category of each sample based on the nearest distance of the neighbors in the feature space. Then, the prediction results are compared with the actual labels to construct a confusion matrix, which shows the number of correct predictions (True Positives and True Negatives) and prediction errors (False Positives and False Negatives) for each category.
Evaluation of the combined Wavelet and GLCM model for texture feature extraction of various types of meat images shows the effectiveness of the integration of these two techniques. Wavelet transform excels in capturing spatial-frequency features, while GLCM effectively represents texture patterns through statistical relationships of pixel intensities.
The research results showed that this hybrid approach significantly improved the ability to distinguish different meat categories (fresh, frozen, and rotten) across different types of meat, including beef, buffalo, lamb, horse, and pork. The combined model achieved an average classification accuracy of 97.2% using the k-NN method, indicating its robustness and reliability in feature extraction and classification tasks.
In addition, based on the analysis results, it was found that the wavelet coefficients provide complementary information with the statistical texture features of GLCM, thereby increasing the separation between classes in the feature space. Evaluation using a confusion matrix confirmed the reliability of the model, with most predictions being in the correct category, although there were some misclassifications due to overlapping feature characteristics in several categories.
For future research, it is recommended to explore the use of more sophisticated classification models, such as Support Vector Machines (SVM), Random Forest, or even deep learning-based approaches such as Convolutional Neural Networks (CNN), which have been shown to be able to identify complex patterns in images with high performance.
[1] Rojas, H.E., Contreras, R.L., Schnake, F.G., Marín, P.M., Huissier, P.G.L., Castillo, P.B. (2022). Factors determining meat quality and cold preservation methods to extend shelf life. Open Access Journal of Biomedical Science, 4(1): 1459-1470.
[2] Cardenas, F.C., Olivera, D.F. (2016). Texture changes in meat during storage. Reference Module in Food Sciences. https://doi.org/10.1016/B978-0-08-100596-5.03294-7
[3] Devi, R., Rasane, P., Kaur, S., Singh, J. (2019). Meat and neat losses: Influence on meat quality. International Journal of Research and Analytical Reviews (IJRAR), 6(1): 762-786. http://www.ijrar.org/papers/IJRAR1BWP104.pdf.
[4] Shi, Y., Wang, X., Borhan, M.S., Young, J., Newman, D., Berg, E., Sun, X. (2021). A review on meat quality evaluation methods based on non-destructive computer vision and artificial intelligence technologies. Food Science of Animal Resources, 41(4): 563-588. https://doi.org/10.5851/kosfa.2021.e25
[5] Taheri-Garavand, A., Fatahi, S., Omid, M., Makino, Y. (2019). Meat quality evaluation based on computer vision technique: A review. Meat Science, 156: 183-195. https://doi.org/10.1016/j.meatsci.2019.06.002
[6] Hidayat, T., Saputri, D.U.E., Aziz, F. (2022). Meat image classification using deep learning with Resnet152V2 architecture. Jurnal Techno Nusa Mandiri, 19(2): 131-140. https://doi.org/10.33480/techno.v19i2.3932
[7] Farinda, R., Firmansyah, Z.R., Sulton, C., Wijaya, I.S., Bimantoro, F. (2018). Beef quality classification based on texture and color features using SVM classifier. Journal of Telematics and Informatics (JTI), 6(3): 201-213.
[8] Priyatno, A.M., Putra, F.M., Cholidhazia, P., Ningsih, L. (2020). Combination of extraction features based on texture and colour feature for beef and pork classification. Journal of Physics: Conference Series, 1563(1): 012007. https://doi.org/10.1088/1742-6596/1563/1/012007
[9] Liu, J., Liu, S.L., Medhat, M.E., Elsayed, A.M.M. (2023). Wavelet transform theory: The mathematical principles of wavelet transform in gamma spectroscopy. Radiation Physics and Chemistry, 203: 110592. https://doi.org/10.1016/j.radphyschem.2022.110592
[10] Ganesh, S.S., Mohanaprasad, K., Karuna, Y. (2016). Object identification using wavelet transform. Indian Journal of Science and Technology, 9(5): 1-7. https://doi.org/10.17485/ijst/2016/v9i5/87187
[11] Fortuna-Cervantes, J.M., Ramírez-Torres, M.T., Mejia-Carlos, M., Murguía, J.S., Martinez-Carranza, J., Soubervielle-Montalvo, C., Guerra-García, C.A. (2022). Texture and materials image classification based on wavelet pooling layer in CNN. Applied Sciences, 12(7): 3592. https://doi.org/10.3390/app12073592
[12] Salman, A.H., Al-Jawher, W. (2024). Enhancing document image classification with discrete wavelet transform and swin transformer. Mathematical Modelling of Engineering Problems, 11(10): 2809-2816. https://doi.org/10.18280/mmep.111023
[13] Widiyanto, S., Sukra, Y., Madenda, S., Wardani, D.T., Wibowo, E.P. (2018). Texture feature extraction based on GLCM and DWT for beef tenderness classification. In 2018 Third International Conference on Informatics and Computing (ICIC), Palembang, Indonesia, pp. 1-4. https://doi.org/10.1109/IAC.2018.8780569
[14] Guo, T., Zhang, T., Lim, E., Lopez-Benitez, M., Ma, F., Yu, L. (2022). A review of wavelet analysis and its applications: Challenges and opportunities. IEEE Access, 10: 58869-58903. https://doi.org/10.1109/ACCESS.2022.3179517
[15] Widianti, L.W., Sudiro, S.A., Madenda, S., Harlan, J. (2020). Cropping method on grayscale images for periapical radiographs of human teeth. In IOP Conference Series: Materials Science and Engineering, 879(1): 012114. https://doi.org/10.1088/1757-899X/879/1/012114
[16] Mutlag, W.K., Ali, S.K., Aydam, Z.M., Taher, B.H. (2020). Feature extraction methods: A review. Journal of Physics: Conference Series, 1591(1): 012028. https://doi.org/10.1088/1742-6596/1591/1/012028
[17] Yang, Y., Park, D.S., Huang, S., Rao, N. (2010). Medical image fusion via an effective wavelet-based approach. EURASIP Journal on Advances in Signal Processing, 2010: 1-13. https://doi.org/10.1155/2010/579341
[18] Aina, T.S., Akinte, O.O., Ademola, B., Iyaomolere, Abode, I.R. (2021). Wavelet transforms and image approximation based image compression system. International Journal of Scientific & Technology Research, 10(10): 2-5.
[19] Vyas, R., Kanumuri, T., Sheoran, G., Dubey, P. (2017). Co-occurrence features and neural network classification approach for iris recognition. In 2017 Fourth International Conference on Image Information Processing (ICIIP), Shimla, India, IEEE, pp. 1-6. https://doi.org/10.1109/ICIIP.2017.8313730
[20] Girisha, A.B., Chandrashekhar, M.C., Kurian, M.Z. (2013). Texture feature extraction of video frames using GLCM. International Journal of Engineering Trends and Technology (IJETT), 4(6): 2718-2721.
[21] Kurniati, F.T., Manongga, D.H., Sediyono, E., Prasetyo, S.Y.J., Huizen, R.R. (2024). GLCM-Based feature combination for extraction model optimization in object detection using machine learning. arXiv Preprint arXiv: 2404.04578. https://doi.org/10.26555/jiteki.v9i4.27842
[22] Deng, J., Bei, S., Shaojing, S., Zhen, Z. (2020). Feature fusion methods in deep-learning generic object detection: A survey. In 2020 IEEE 9th Joint International Information Technology and Artificial Intelligence Conference (ITAIC), Chongqing, China, pp. 431-437. https://doi.org/10.1109/ITAIC49862.2020.9338904
[23] Sudha, D., Ramakrishna, M. (2017). Comparative study of features fusion techniques. In 2017 International Conference on Recent Advances in Electronics and Communication Technology (ICRAECT), Bangalore, India, pp. 235-239. https://doi.org/10.1109/ICRAECT.2017.39
[24] Subbu, R.S., (2017). Brief study of classification algorithms in machine learning approved. Master’s Thesis, City University of New York.
[25] Mohanapriya, M., Lekha, J. (2018). Comparative study between decision tree and KNN of data mining classification technique. Journal of Physics: Conference Series, 1142(1): 012011. https://doi.org/10.1088/1742-6596/1142/1/012011
[26] Shokrzade, A., Ramezani, M., Tab, F.A., Mohammad, M.A. (2021). A novel extreme learning machine based kNN classification method for dealing with big data. Expert Systems with Applications, 183: 115293. https://doi.org/10.1016/j.eswa.2021.115293
[27] Varoquaux, G., Colliot, O. (2023). Evaluating machine learning models and their diagnostic value. Machine Learning for Brain Disorders, Humana, New York, NY. 601-630. https://doi.org/10.1007/978-1-0716-3195-9_20
[28] Xu, Y., Goodacre, R. (2018). On splitting training and validation set: A comparative study of cross-validation, bootstrap and systematic sampling for estimating the generalization performance of supervised learning. Journal of Analysis and Testing, 2(3): 249-262. https://doi.org/10.1007/s41664-018-0068-2
[29] Hall-Beyer, M. (2017). Practical guidelines for choosing GLCM textures to use in landscape classification tasks over a range of moderate spatial scales. International Journal of Remote Sensing, 38(5): 1312-1338. https://doi.org/10.1080/01431161.2016.1278314
[30] Yildirim, M., Cinar, A. (2020). Classification of Alzheimer's disease MRI images with CNN based hybrid method. Ingénierie des Systèmes d’Information, 25(4): 413-418. https://doi.org/10.18280/isi.250402
[31] Sathyanarayanan, S., Tantri, B.R. (2024). Confusion matrix-Based performance evaluation metrics. African Journal of Biomedical Research, 4023-4031. https://doi.org/10.53555/AJBR.v27i4S.4345
