Robust Algorithm for Deep Face Recognition System Performance Enhancement Based on Hybrid Techniques

Abstract

The effectiveness of most face recognition systems would be greatly reduced if the images in the dataset had inconsistent lighting conditions in uncontrolled environments. To address this problem, a new algorithm was developed that combines Modified Difference of Gaussian (MDoG) with Discrete Wavelet Transformation (DWT) and various edge detection operators, including first-order derivative filters (such as Sobel, Prewitt, and Roberts filters) and second-order derivative filters (such as Zero cross, LoG, and Canny filters). A features extractor based on Linear Regression Slope (LRS) and Principal Components Analysis (PCA), was used to extract features from the images. The accuracy of the algorithm was conducted with the Optimized Artificial Neural Network (OANN) for classification purpose and with different datasets including CNF, CK, JAFFE, and CAS. The experiment proves that this algorithm has greater efficiency than other modern methods, the best results are obtained with  MDoG-DWT-SF, LRS, and OANN pairing and also MDoG-DWT-ZF, LRS, and OANN combination.