Design and Enhancement of a CNN Model to Augment the Face Recognition Accuracy

In the last decade, smart home security applications have relied more on human biometrics in their functions, due to the reliability and the high-precision results these technologies provide. Face recognition is one of the popular biometrics in the field of image processing technologies. Human face recognition processing is a complicated operation that involves different factors and circumstances such as the illumination degree and the position of the face that affects the final recognition rate. In this research, the Convolution Neural Network (CNN) architecture is used in the extracting phase of significant features of the face shape, and the SoftMax classification layer was used to identify faces in the fully connected CNN layer. This paper provides an update of CNN architecture by applying a three-batch normalization layer to the CNN design. By applying this modification, the system network speed increased with a better recognition rate. The recognition rate also increased by applying two DWT levels with a bio5.5 filter to the training group of the database images and the tested image before applying the PCA dimensional reduction algorithm instead of using the PCA algorithm alone. The obtained face recognition rates have been improved to 99.75% by applying the proposed CNN scheme. While applying the proposed hybrid approach (using the PCA next to applying DWT-2 levels with bior5.5 filter) has registered a 99.25% recognition rate compared to a 96.75% recognition rate when obtained by applying the PCA method alone. The research has adopted using a set of 360 training images and 40 test images set of the standard ORL Database in its work.