A proposed 3-stage CNN classification model based on augmentation and denoising

This work proposed a CNN classification model that aims to classify the faces by three stages applied to a real data set. The first stage shows the effects of the augmentation technique on the real data set where these effects include online, offline, and without augmentation. At this stage, the proposed CNN model is a built-from-scratch that has low computational complexity, low layers and the smallest filter sizes. The second stage involved denoising the images in the real data set, where the images are preprocessed by applying the median, Gaussian, and mean filters to render the images more smooth and compare the effects of these filters based on the classification accuracy. The third stage involved a multi-class proposed model that contained 12 classes of images that were trained on the applied real data set, in addition to a benchmark set of images that was collected from the Internet. The findings reveal that the model accuracy reached 98.81% when the offline augmentation model or the median filter was applied to the real data set, while it reached 97.48% when the CNN multi-class proposed model was applied to identify the non-permission class. These processes were found to improve the performance parameters such as precision, recall, F1 score, and area under the curve (AUC). Finally, to enhance the test prediction accuracy and test time, pre-training and fine-tuning (transfer learning) are applied on the real data set so as test accuracy and test time of our proposed model are better as compared with other models reached 99.7% and 4 seconds respectively.