A modified retinex illumination normalization approach for infant pain recognition system

Abstract

Pains in newborn babies are monitored in a Neonatal Intensive Care Unit (NICU) for medical treatment. Pain in newborns can be detected by studying their facial appearance. Even though the outcome is acceptable, it is not adequately vigorous to be used in unpredictable, non-ideal situations such as noise and varying illumination environment. First, to improve the noise cancellation robustness an adaptive median filter (AMF) is proposed. Mean and variance of median values are selected to generate a weight for each window part of the images such as 3x3, 5x5 or 7x7. Various linear and nonlinear filters are adopted to eliminate the noise in the images. Quantitative comparisons are performed between these filters with our AMF in terms of Peak Signal-to-Noise Ratio (PSNR), Mean Square Error (MSE), Image Enhancement Factor (IEF) and Mean Structural SIMilarity (MSSIM) Index. The average results show improvement in terms of 40.63 db for PSNR, 6.01 for MSE, 258.09 for IEF and 0.97 for MSSIM respectively. In this work a novel method of illumination invariant normalization known as Modified Retinex Normalization (MRT) for preprocessing of infant face recognition is proposed. This is based on a modified retinex model that combines with histogram normalization for filtering the illumination invariant. The proposed method is compared to other methods like Single scale Retinex (SSR), Homomorphic method (HOMO), Single Scale Self Quotient Image (SSQ), Gross and Brajovic Technique (GBT), DCT-Based Normalization (DCT), Gradientfaces-based normalization technique (GRF), Tan and Triggs normalization technique (TT), and Large-and small-scale features normalization technique (LSSF) for evaluation with Infant Classification of Pain Expressions (COPE) database. Several experiments were performed on COPE databases. Single PCA, LBP and DCT feature extraction information yielded a good recognition result. However, by summing these three, it gives more robustness to noise and illumination classification rate because the sum rule was the most resilient to estimate errors and gives higher than 90% accuracies of pain and no pain detection. The new illumination normalization and combination of features gives higher results of more than 90% on five different classifiers with various algorithms such as k-nearest neighbors (k-NN), Fuzzy k-nearest neighbors (FkNN), Linear Discriminat Analysis (LDA), Feed Forward Neural Network (FFNN), Probabilistic Neural Network (PNN), General regression Neural Network (GRNN), SVM Linear kernel (SVMLIN), SVM RBF kernel (SVMRBF), SVM MLP kernel (SVMMLP) and SVM Polynomial kernel (SVMPOL) with different performance measurement such as Sensitivity, Specificity, Accuracy, Area under Curve (AUC), Cohen's kappa (k), Precession , F-Measure and Time Consumption .