Due to the wide range of applications of emotions in understanding human behavior, diagnosing mental-related disease, human-computer interaction, and robotics, emotion recognition becomes an emerging research domain. Human facial emotions are used to deliver their feelings to other people, and also it supports to know the intentions of people. Recently, several methods have been developed to recognize emotions based on the input data, but they face issues, such as high computational cost, lower performance, more training time, and low-quality data. To overcome these problems, Optimization based Self-supervised enabled Distributed Bidirectional Model (OSDB) is proposed in this research. The model takes multimodal data like video, audio, and text as input to identify the facial emotions accurately. To select the features that, in turn, enable the model to provide better results, the Levy Migration Optimization (LeMO) is specifically used. In the research, the utilization of a self-supervised learning mechanism increases the generalization ability and reduces the computational complexity. Thus, the proposed OSDB model reaches a performance with respect to accuracy as 98.7%, Mathews Correlation Coefficient (MCC) as 0.96 %, F1-Score as 98.7%, specificity as 98.9%, and sensitivity as 98.4% in 90% of the training data.