Abstarct: Abstract In recent years, the Abstraction and Decision Fusion Architecture (ADFA) has been developed to balance accuracy, processing speed, and efficient resource utilization in machine learning models. This architecture has been applied in tasks such as sign language recognition and handwritten digit classification. ADFA is structured into three distinct laxyers: the abstraction laxyer, the computation laxyer, and the fusion laxyer. In the abstraction laxyer, data summarization routines are developed. In the computation laxyer, lightweight computational models are designed and implemented, each receiving the output of one abstraction routine as input. In the fusion laxyer, the decisions of the models in the second laxyer are fused to produce a final unified decision. In this thesis, after investigating the ADFA architecture, we adapt and apply it to develop machine learning models in two domains: Persian handwritten character classification and emotion recognition. In the handwriting classification task, the proposed model maintains high accuracy and demonstrates strong computational efficiency and resource-conscious performance, achieving an accuracy of over 96% on the IFHCDB image dataset and outperforming existing state-of-the-art models in terms of accuracy, efficiency, and resource usage. In the domain of emotion recognition, the proposed models achieve over 96% accuracy on the L-SVD video dataset and, compared to leading models in this field, deliver higher accuracy and significant reductions in computational load and memory usage. The results of this research highlight the strong potential of the ADFA architecture for developing machine learning models in image and video processing applications.