الفهرس | Only 14 pages are availabe for public view |
Abstract In this thesis, two different signal processing methods with SVMs in classifying infant cry signals are examined to determine the best accuracy among the two techniques. The cry signals are obtained from Baby Chillanto database , then passed into feature extraction techniques .The two signal processing techniques used are either MFCC, or DWT followed by Mel frequency feature extraction technique. Both features enter a SVM for classifying normal , asphyxia and deaf cry signals. To assess the performance of our approach , seven statistical measures are used. 5.2 Conclusion By applying linear and RBF kernel with different sigma values SVMs with both validation schemes in classification stage, we examined the classification accuracy of normal versus asphyxia cry signals experiment, and conducted that decomposing the input cry signal into approximated and detailed coefficients followed by estimation of the MFCC from decomposed signals gives more promising results than directly extracting the cepstrum features from input cry signal. The best classification accuracy obtained was 97.8 % at MFCC value equals to 6 using RBF kernel with (σ =12) with cross validation. In classifying normal and deaf cry signals , MFCC technique performed better than DWT-MFCC and the best accuracy obtained was 99.9% using RBF kernel(ơ=12) with cross validation at number of MFCC coefficient C = 10. 5.3 Recommendation It is recommended to use the proposed system with different machine learning algorithms for detecting other types of neonatal pathologies. Also, our future goal is to apply the two signal processing techniques on real infant cry signals dataset. |