Utilizing Deep Learning Models to Identify and Classify Various Types of Infant Cries. رسالة ماجستير

Abstract

Current research emphasizes the development of effective models to aid specialists in the efficient classification of behavioral conditions, one of these fields is infant crying. Infant crying is a critical indicator of a child's health, functioning as a non-verbal communication method that conveys their needs, understanding the cause of the cry of an infant could be quite a challenging task, especially for parents and caregivers. Therefore, this problem becomes an area of research for researchers. Deep learning utilization is increased in healthcare, such as private clinics and hospitals. The utilization of deep learning methodologies can improve earlier diagnosis; it enhances their ability to diagnose the causes of the infant crying without subjecting the infant to greater suffering. To those seeking to improve diagnostic tools for classification of infant cries, this study provides important insights into the factors of cry features that correlate lightly with cry function and function condition. The capability of deep learning techniques applied in this work for boosting diagnostic precision and results in pediatric healthcare is checked. Techniques for converting audio to images using MFCC has been employed and then classified the resulting images with different deep-learning techniques (VGG16, DenseNet, IV ResNet, and GoogLeNet). Appropriate matrices techniques were used to evaluate and compare the performance of these models. Deep learning techniques have been examined in the classification of infant cries, utilizing data gathered from local care and global dataset platforms such as Kaggle. And then, a comprehensive analysis of the performance of various algorithms to identify the most suitable approaches for accurately classifying infant cries in diverse data environments is performed. According to the result, ResNet and DenseNet performed better than the others, with ResNet reaching higher accuracy on global dataset and DenseNet following. The effectiveness of VGG-16, on the contrary, was lowest, reflecting that each algorithm must be evaluated on its own effectiveness relative to the dataset it is used with. With local data, ResNet has high accuracy and proves its reliability and robustness on this type of data. in addition, the study highlights the importance of architectural improvements in enhancing model performance. The incorporation of a Random Forest (RFF) layer with Dense layers resulted in significant enhancements in model stability and accuracy. The architectural modifications suggest that hybrid approaches can significantly improve the model's capacity to learn complex patterns, which is essential for applications in medical imaging and diagnostics. By the above, it follows that we can aid service providers in recognizing accurately and fast the real cause of an infant’s cry using the latest deep learning techniques. Also, A model can be generated and distributed to these centers of the future to benefit from it.