A Novel Classification Layer Technique for ASL Analysis Using CNN

American sign language (ASL) is the most well-known sign language in deaf communities and has expanded to be a sign language in other countries, including Thai sign language (TSL). Currently, computer vision and artificial intelligence (AI) technologies are fast stepping and expanding into many fields of human living. There is much knowledge in ASL applications, which can also be exploited to develop TSL applications in the future. This study will analyze the Convolution Neural Network classifier (CNN) type for ASL, ranging from ordinary CNN classifiers, CNN with Support Vector Machine (SVM) classifiers, and Ensemble CNN classifiers. We set the augmented validation dataset to mimic the natural world environment. The new CNN model is better than the previous state of art ASL CNN model in the augmented validation dataset. Furthermore, to improve the proposed CNN model, we tested 2 experiments, first was to replace the dense classifier in the CNN model with an SVM classifier, and the second was to ensemble two pieces of knowledge from 2 learned CNN models with averaging aggregation of SoftMax's probability. Finally, the best validation accuracy can be achieved by a novel model from the CNNs' ensemble techniques with averaging aggregation of SoftMax's probability.