Analysis of Denoising Techniques applied to Facial Images for Emotion Recognition

The recognition accuracy of automated facial emotion recognition (AFER) is affected by many factors i.e. emotion intensity, ageing, facial hair, image resolution, head pose, presence of noise etc. Thus preprocessing stage in AFER is quite crucial and necessitates the proper selection of filter. Therefore primary aim of this work is to remove noise from facial images using Savitzky-Golay filter (SGF), Discrete Wavelet Transform (DWT) and median filters. The filters are implemented on images taken from extended Cohn-Kanade (CK+) database, corrupted by various noises with varying noise levels. Results reveal the effectiveness of designed filters for AFER in terms of coefficient of correlation (COC), signal to noise ratio (SNR), mean square error (MSE), peak signal to noise ratio (PSNR) and structure similarity (SSIM).