Generalized regression neural network trained preprocessing of frequency domain correlation filter for improved face recognition and its optical implementation

The paper proposes an improved strategy for face recognition using correlation filter under varying lighting conditions and occlusion where spatial domain preprocessing is carried out by two convolution kernels. The first convolution kernel is a contour kernel for emphasizing high frequency components of face image and the other kernel is a smoothing kernel used for minimization of noise those may arise due to preprocessing. The convolution kernels are obtained by training a generalized regression neural network using enhanced face features. Face features are enhanced by conventional principal component analysis. The proposed method reduces the false acceptance rate and false rejection rate in comparison to other standard correlation filtering techniques. Moreover, the processing is fast when compared to the existing illumination normalization techniques. A scheme of hardware implementation of all optical correlation technique is also suggested based on single spatial light modulator in a beam folding architecture. Two benchmark databases YaleB and PIE are used for performance verification of the proposed scheme and the improved results are obtained for both illumination variations and occlusions in test face images.

► Modification of correlation filter with trained preprocessing is proposed.
► Spatial preprocessing is done by using two convolution kernels.
► Kernel elements are generated using generalized regression neural network.
► Required Fourier transforms are performed in an all-optical architecture.
► The proposed technique shows improved performance during face recognition.