Non-linear features and multi-objective based feature selection algorithms for infant cry classification

Automatic infant cry recognition has become popular research in the past decade and this scenario may help in early detection of infant health status. The infant cry classification is classifying the health status of infant based on their cry signals. Two databases were utilized, where first database contains 340 sample of Asphyxia, 507 sample of Normal, 879 samples of Deaf, 350 samples of Hungry and 192 sample of Pain. Another database contains 45 sample of Normal, 531 sample of Premature, and 513 sample of Jaundice. Linear features such as Mel-Frequency Cepstral Coefficient (MFCC) and Linear Predication Coefficient (LPC) are commonly applied in extracting cry signals that only relies on frequency based information, thus, time-frequency domain feature extraction methods named Dual Tree Complex Wavelet Transform (DTC-WT) and Dual Tree Complex Wavelet Packet Transform (DTC-WPT) are employed. A total of 2176 features were extracted from the overall methodology for the analysis. The huge number of features has a tendency to complicate the computation process and reduce accuracy and often cited in literature as curse of dimensionality. In order to reduce the high dimensionality feature dataset, wrapper based feature selection techniques was proposed to select the most relevant features. Two important outputs concerned during feature selection stage are number of selected features and classification accuracy. Therefore, multi-objective optimization algorithms are proposed in optimizing both objectives simultaneously during reducing size of feature dataset. Due to the there is no particular crossover and mutation technique is universally solving all MOP, also duplication of solutions during NSGA-II optimization which cause the difficulty of searching the optimal solutions. A novel Modified Non-Dominated Sorting Genetic Algorithm–II (MNSGA-II) was proposed by adopting the population updating mechanism into conventional NSGA-II by replacing crossover and mutation technique by sine cosine functions from Sine Cosine Algorithm (SCA). Kursawe and Zitzler-Deb-Thiele (ZDT) test functions were apply to validate the performance of proposed MNSGA-II. MNSGA-II able to achieved lowest generational distance (GD) of 1.22e-03, 1.01e-04, 3.59e-05,and 1.23e-04 for Kursawe, ZDT1, ZDT2, and ZDT3 respectively which indicated that the optimal solutions found are near to true Pareto front. Finally MNSGA-II wrapper based feature selection technique was applied to solve five class and seven class infant cry classification problems. Another medical database (microarray cancer gene classification) with huge dimensionality was also used to validate the robustness of proposed MNSGA-II. Several binary infant cry classification were conducted to examine the performance of DTC-WT and DTC-WPT features, maximum accuracy of 99.85% and 100% for Deaf vs Normal were achieved by DTC-WT and DTC-WPT respectively. MNSGA-II reduces the feature dimension for five classes (Hunger, Deaf, Normal, Asphyxia and Pain) with 511 features and accuracy of 96.82%, while 573 features and accuracy of 96.87% for seven classes classification (Hunger, Deaf, Normal, Asphyxia, Pain, Jaundice and Premature). Finally the overall classification results from microarray cancer gene classification proven the robustness of MNSGA-II wrapper based feature selection in pattern classification.