<p>Understanding the emotions of a person is an important pillar of learning human psychology. Various psychological studies indicate that non-verbal communication is 55% parts of the overall human behaviour. It provides a scope to analyse human emotions based on these non-verbal cues. Numerous, machine learning and deep learning methodologies can be used for emotion detection. In this study, a deep learning based hyperparameter tuned model is proposed for human emotion detection. The proposed model is tested and validated on various benchmark datasets named FER-2013, CK + and JAFFE using state of the art models like MobileNetV2, InceptionNet and AttentionCNN. Two augmented datasets namely Ensemble and augmented ensemble are proposed to validate the emotion recognition. These curated datasets are based on the benchmark datasets available for emotion detection. Experiments prove that on the devised datasets the proposed model detects multi-class emotions in real time. The proposed E_image model demonstrates superior performance in emotion recognition tasks, achieving a validation accuracy of 68.53% and lower validation loss of 0.959 on the Ensemble dataset. This represents an improvement of validation accuracy approximately 6% and betterment in validation loss over MobileNetV2 (62.02%, 1.676) and over 12% higher than AttentionCNN (56.27%), indicating substantial advancement over existing models. Numerous loss functions are considered for the statistical validity of the proposed model.</p>