Exploring Machine Learning for Image Enhancement and Multimedia Understanding towards Autistic People’s behavior analysis

This dissertation investigates the application of machine learning techniques for monitoring and analyzing signs and behaviors associated with Autism Spectrum Disorder (ASD), with a particular emphasis on image and audio processing. From an Explainable Artificial Intelligence perspective, the research begins by examining the capabilities of deep learning models in the analysis of facial features of autistic and non-autistic individuals. A crucial element in this process is the enhancement of input image quality. To address this, we also propose a multi-exposure High Dynamic Range (HDR) imaging method, which improves image detail through segmentation and deep learning-based reconstruction. The proposed method can be used for various topics, one of which is for our research on distinguishing autistic from non-autistic individuals. The IEEE ICASSPW paper highlights the superior performance of empirical thresholding over Otsu thresholding; however, the integration of these methods led to overfitting, prompting the adoption of Otsu segmentation. The results of the reconstruction network, consisting of Visual Attention Modules (VAM), attention and alignment modules, and refinement stages, outperformed state-of-the-art techniques, as published in IEEE Access. Additionally, the research explores the use of Vision Transformers and ResNets to differentiate children with autism from their neurotypical peers, achieving a 92% accuracy rate. We also use explainable AI techniques to clarify the model's decision-making process, with findings submitted to the Journal of Research in Autism Spectrum Disorders. Furthermore, the study investigates vocal stereotypy measurement in autistic children using machine learning applied to audio files, resulting in tailored models for individual patients. These findings were submitted to the Journal of Applied Behavior Analysis.