An Effective Face Recognition Method Using a Guided Image Filter and a Convolutional Neural Network

Abstract

Face recognition remains difficult in computer vision because pose, facial expression and illumination all vary, and 
performance drops in unconstrained environments. This paper implements a face recognition method combining a 
guided image filter with a convolutional neural network. The face region is first located in the input image using the 
Viola-Jones detector and resized to a fixed size, then passed through a guided image filter. The guided filter is an 
edge-preserving smoothing operator, which is the property that matters here: it suppresses the noise and fine texture 
variation that differ between images of the same person while leaving the facial edges that carry identity intact, so the 
network is presented with a more consistent input than the raw image provides. A convolutional network of five blocks 
then extracts features and classifies the face. The first four blocks each contain convolution, batch normalization, 
ReLU and max pooling layers, with 5 × 5 filter kernels and 16, 32, 64 and 128 channels respectively, followed by two 
dense layers and a softmax classifier. Experiments on the ORL, JAFFE and YALE face databases attained recognition 
rates of 98.33%, 99.53% and 98.65% respectively. The softmax classifier gave better results than decision tree and 
random forest alternatives in the classifier section of the network.