Over the past two decades, deep learning has shown to be the solution to numerous machine learning issues. It does have two major drawbacks, though: the need for large amounts of labelled data and the expense of training. By applying previously acquired information from a source data or task to a target data or task during training, a technique known as Deep Transfer Learning (DTL) in deep learning aims to lessen this dependency and associated expenses. Identifying and deciphering emotions from facial expressions is the technique of facial emotion recognition. Including as computer science, psychology, and neuroscience. The fields of marketing, security, healthcare, and human-computer interaction will all be impacted by the capacity to precisely identify emotions from facial expressions. With a focus on facial expression theories, facial emotion recognition types, datasets, algorithms, assessment metrics, and applications, this paper provides a comprehensive summary of the state of the art in face emotion detection today. The study also discusses moral issues surrounding the identification of facial emotions. This survey’s objective is to offer a thorough overview of the state of face expression recognition technology as it stands today. The work focuses on employing convolution neural networks (CNNs) to extract features and categorise using the transfer learning technique on pre-trained models. The majority of research perform better on controlled datasets but falter on datasets containing image changes. A deep learning model based on Xception and Convolution Neural Network (CNN) [1], has been developed in recent work to improve emotion recognition. This system can concurrently conduct face recognition and emotion categorization, focusing on key components such as the face. Using a visualisation technique, various emotions are distinguished from one another according to the classifier results. The Xception algorithm successfully completed all tasks, including emotion recognition and classification, with an accuracy of almost 99.99%, according to experimental research conducted on the CK+ dataset.