Academia.eduAcademia.edu

Figure 4 – uploaded by Sanusi AbuDarma

See full PDFdownloadDownload figure
Fig. 4 Shows the sample face images from the Caltech 101_ObjectC atogary database The data acquisition process concerns the methods used to obtain a large number of training and testing datasets for the proposed model. In this paper, the study proposed collecting a sufficient number of datasets from two publicly available databases, namely the Caltech-101-Object-Category database and the Label Face in the Wild (LFW) Database (shown in Figure 4). Caltech's-101-Object-Category database has 450 facial photos of 27 different persons. The photographs are 325 x 495 pixels in Jpeg format, with varying expressions, backgrounds, and lighting, but this study advocated cropping and reducing each face image to 64x64 pixels.

Figure 4 Shows the sample face images from the Caltech 101_ObjectC atogary database The data acquisition process concerns the methods used to obtain a large number of training and testing datasets for the proposed model. In this paper, the study proposed collecting a sufficient number of datasets from two publicly available databases, namely the Caltech-101-Object-Category database and the Label Face in the Wild (LFW) Database (shown in Figure 4). Caltech's-101-Object-Category database has 450 facial photos of 27 different persons. The photographs are 325 x 495 pixels in Jpeg format, with varying expressions, backgrounds, and lighting, but this study advocated cropping and reducing each face image to 64x64 pixels.

Related Figures (21)

tal dice aimee T ‘eaiet ieee aaa Fig 3. Max pooling operation on 4x4 convolutional output to feature map (2x2 windows)
tal dice aimee T ‘eaiet ieee aaa Fig 3. Max pooling operation on 4x4 convolutional output to feature map (2x2 windows)
Fig. 5 Shows the sample face images from label faces in the wild ir resize scale of (64x64) pixels
Fig. 5 Shows the sample face images from label faces in the wild ir resize scale of (64x64) pixels
Fig. 6 is the training progress graph. The first graph at the top represents training accuracy (classification accuracy). The x-axis represents a scale of 0 to 80 epochs and 0 to 880 iterations, respectively, while the y-axis represents an accuracy value scale of 0% to 100%. The loss function (cross- entropy loss errors) is depicted in the bottom graph. This graph shows the iteration used to estimate the gradient. A light blue line and a dark blue line represent the classification accuracy. This accuracy was obtained by implementing a smooth approach to training accuracy.
Fig. 6 is the training progress graph. The first graph at the top represents training accuracy (classification accuracy). The x-axis represents a scale of 0 to 80 epochs and 0 to 880 iterations, respectively, while the y-axis represents an accuracy value scale of 0% to 100%. The loss function (cross- entropy loss errors) is depicted in the bottom graph. This graph shows the iteration used to estimate the gradient. A light blue line and a dark blue line represent the classification accuracy. This accuracy was obtained by implementing a smooth approach to training accuracy.
The proposed model yielded the label count result shown in Table 2 above. The model can figure out how many images are in each class. The label count is a table containing the labels for each class and the number of images [8]. The model is capable of detecting, categorizing, and calculating the number of images in each of the six classes. for proper network training, computational stability, and convergence speed.
The proposed model yielded the label count result shown in Table 2 above. The model can figure out how many images are in each class. The label count is a table containing the labels for each class and the number of images [8]. The model is capable of detecting, categorizing, and calculating the number of images in each of the six classes. for proper network training, computational stability, and convergence speed.
Fig. 6 Training progress plot graph of the proposed 22 layers CNN
Fig. 6 Training progress plot graph of the proposed 22 layers CNN
proposed model has good performance for face recognition applications under different uncontrolled conditions. This means the proposed model is resilient to all six facial image variation classes involving facial expressions, facial makeup, occlusion, age-related variation, and pose.
proposed model has good performance for face recognition applications under different uncontrolled conditions. This means the proposed model is resilient to all six facial image variation classes involving facial expressions, facial makeup, occlusion, age-related variation, and pose.
Table 5. Training plot detail for the 15-layer CNN training from scratch
Table 5. Training plot detail for the 15-layer CNN training from scratch
Where TP, TN, FP, and FN denote true positive, true negative, false positive, and false negative, respectively [11]. To appraise the true performance of the proposed hybrid multilayer CNN+SVM architecture on each of the six classes of face variation, this study adopted an evaluation metric evaluation measures of accuracy, precision, recall, and f1- score to evaluate and prove the good performance of the proposed model on the multi-class classification of six human face image versions. The model accuracy is achieved by dividing the correctly classified samples by all the samples.
Where TP, TN, FP, and FN denote true positive, true negative, false positive, and false negative, respectively [11]. To appraise the true performance of the proposed hybrid multilayer CNN+SVM architecture on each of the six classes of face variation, this study adopted an evaluation metric evaluation measures of accuracy, precision, recall, and f1- score to evaluate and prove the good performance of the proposed model on the multi-class classification of six human face image versions. The model accuracy is achieved by dividing the correctly classified samples by all the samples.
Table 6. Performance evaluation result for 15 layers based on 6 classes of variations and since the two are higher, it means that the model is not making mistakes in classifying the predicted class. However, facial expression, facial makeup, and pose variation have the same accuracy value of 0.9976 and the same precision value of 0.9857 with different recalls and f1-score values. Figure 7 demonstrates that a notable observation may be made at the start of the model's training throughout the first few epochs (1 to 28) and iterations (1 to 300). Both the training and validation accuracies (solid blue and dotted red lines begin with higher volatility due to overfitting. Both the training and validation accuracy plateaued for a short period at epoch 32 until the gradient updates escaped the unfavorable local minimum and began to converge at epoch 41 and iteration 450 with training and validation accuracies of 99.25% and 98.45%, respectively. At epoch 50, the training accuracy dropped to 98.50% until epoch 66, when it restored its typical accuracy value of 99.25%, while at the same epoch 66, the validation accuracy increased to 98.65%.
Table 6. Performance evaluation result for 15 layers based on 6 classes of variations and since the two are higher, it means that the model is not making mistakes in classifying the predicted class. However, facial expression, facial makeup, and pose variation have the same accuracy value of 0.9976 and the same precision value of 0.9857 with different recalls and f1-score values. Figure 7 demonstrates that a notable observation may be made at the start of the model's training throughout the first few epochs (1 to 28) and iterations (1 to 300). Both the training and validation accuracies (solid blue and dotted red lines begin with higher volatility due to overfitting. Both the training and validation accuracy plateaued for a short period at epoch 32 until the gradient updates escaped the unfavorable local minimum and began to converge at epoch 41 and iteration 450 with training and validation accuracies of 99.25% and 98.45%, respectively. At epoch 50, the training accuracy dropped to 98.50% until epoch 66, when it restored its typical accuracy value of 99.25%, while at the same epoch 66, the validation accuracy increased to 98.65%.
Table 7. Training plot details for the 9 layers CNN+SVM training from scratch
Table 7. Training plot details for the 9 layers CNN+SVM training from scratch
Table 8. Performance evaluation result for 9 layers CNN+SVM on the classes of variation
Table 8. Performance evaluation result for 9 layers CNN+SVM on the classes of variation
Table 9. Training Plot Details for 7 Layers CNN&SVM, Training from Scratch
Table 9. Training Plot Details for 7 Layers CNN&SVM, Training from Scratch
Fig. 10 Training and validation losses (Cross entropy curve) for 9 LYCNN
Fig. 10 Training and validation losses (Cross entropy curve) for 9 LYCNN
Fig. 11 Training and validation accuracies (Classification accuracies curve) for 7-layer CNN
Fig. 11 Training and validation accuracies (Classification accuracies curve) for 7-layer CNN
Fig. 12 Training and validation losses (Cross entropy curve) for 7 LYCNN
Fig. 12 Training and validation losses (Cross entropy curve) for 7 LYCNN
Fig. 13 Validation loss comparison between the proposed MLTYLY CNN and three deep learning model architectures
Fig. 13 Validation loss comparison between the proposed MLTYLY CNN and three deep learning model architectures
model is unstable and has a higher overfitting problem. Generally, it can be stated that the proposed hybrid model performs well in classifying all six classes of face variations. 9.5. Comparative Analysis of the F our Models Performance Evaluation Results
model is unstable and has a higher overfitting problem. Generally, it can be stated that the proposed hybrid model performs well in classifying all six classes of face variations. 9.5. Comparative Analysis of the F our Models Performance Evaluation Results
Related topics:
Artificial IntelligenceComputer VisionMachine Learning
580 California St., Suite 400
San Francisco, CA, 94104
© 2025 Academia. All rights reserved